Rebooting Ultrasonic Positioning Systems forUltrasound-incapable Smart Devices

Qiongzheng Lindagger Zhenlin Andagger Lei YangdaggerCo-primary Authors

Department of ComputingThe Hong Kong Polytechnic University

linanyoungtagsysorg

ABSTRACTAn ultrasonic Positioning System (UPS) has outperformedRF-based systems in terms of its accuracy for years How-ever few of the developed solutions have been deployed inpractice to satisfy the localization demand of todayrsquos smartdevices which lack ultrasonic sensors and were consideredas being ldquodeafrdquo to ultrasound A recent finding demonstratesthat ultrasound may be audible to the smart devices undercertain conditions due to their microphonersquos nonlinearityInspired by this insight this work revisits the ultrasonic posi-tioning technique and builds a practical UPS calledUPS+ forultrasound-incapable smart devices The core concept is todeploy two types of indoor beacon devices which will adver-tise ultrasonic beacons at two different ultrasonic frequenciesrespectively Their superimposed beacons are shifted to alow-frequency by virtue of the nonlinearity effect at the re-ceiverrsquos microphone This underlying property functions asan implicit ultrasonic downconverter without throwing harmto the hearing system of humans We demonstrate UPS+ afully functional UPS prototype with centimeter-level local-ization accuracy using custom-made beacon hardware andwell-designed algorithms

CCS CONCEPTSbull NetworksrarrMobile networks

KEYWORDSUltrasonic Positioning System Nonlinearity Effect MobileComputing Smart Devices UPS+

1 INTRODUCTIONTracking smart devices such as phones or wearables insidebuildings where the GPS is not available has become a grow-ing business interest Indoor localization enables users tonavigate indoor spaces similar to the function provided byGPS for outdoor environments It prompts a series of keymobile applications namely indoor navigation (eg malls

cBeacon

uBeaconuBeacon

uBeacon

Fig 1 System architectureMultiple uBeacons are deployed aslocation anchors for the trilateration whereas a single cBeacon isinstalled for ultrasonic downconversion

factories and airports) augmented reality location-awarepervasive computing advertising and social networking

Many efforts have been exerted to deliver an accuracy oftens of submeters for indoor localization However only afew of these solutions have actually reached todayrsquos mobiledevices because previous works have failed in one of twocategories at a high level that is they either require mobiledevices to be equipped with specialized sensors such as ac-celerators [1] magnetic sensors [2] LEDs [3 4] RFIDs [5ndash9]and WiFi [10] etc or they require exhaustive fingerprintingof the environment to learn the spatial distribution of signalcharacteristics such as FM [11] WiFi [12 13] Bluetooth [14]sound or light [15 16] and Zigbee [17] etc Ultra-wide band(UWB) (eg WiTrack [18]) can achieve cm- (even mm-) levelaccuracy but requires GHz bandwidth which is not allowedin practice due to the spectrum regulations On the contraryas the rapid development of smart technology The demandfor highly accurate indoor localization services has becomea key prerequisite in some markets thereby resulting in agrowing business interest For example finding the AirPodin a head tracking userrsquos arms writing in the air pinpointingitems in VRAR systems and so on

In this work we revisit a classical indoor localization solu-tion namely the ultrasonic positioning system (UPS) which

arX

iv1

812

0234

9v2

[cs

NI]

10

Dec

201

8

cBeacon uBeacon

WiFi Chipset

Transducer (recv)

Transducer (trans)Beacon devices(a) (b) (c)

Fig 2 UPS+ experimental platform (a) Experimental scenario (b) two beacon devices cBeacon and uBeacon and (c) zoom-in image ofthe circuit board of the uBeacon

utilizes the ultrasonic sound as the ranging media UPS out-performs RF-based systems in terms of accuracy For exampleMIT launched Project Oxygen [19] in 2004 In this project apioneering UPS named Cricket [20] was developed to providecentimeter-level accuracy location service Subsequently alarge number of follow-up works [21ndash32] have further uti-lized sound or ultrasound for ranging or localization (seesect2) However despite the extensive efforts exerted on UPSresearch only a few of the resulting solutions have been ac-tually deployed in practice at present For example Cricketproject ceased updating 10 years ago One of the importantreasons why UPSs do not receive enough attention today isthat they suffer from a serious defect that is todayrsquos smartdevices lack ultrasonic sensors and therefore cannot receiveultrasonic beacons from UPSs [25 28 31] To bridge this gapmany follow-up works [25 28 31] have advised transmittingbeacons at the spectrum of 20 sim 22 kHz where 22 kHz isthe upper bound of currently available microphones Thesound within this band is supposedly inaudible to the major-ity of humans Nevertheless it is still harmful to infants andpets who are hyper-sensitive to the high-frequency soundsMoreover the signals near the upper bound may be seriouslydistorted and attenuated due to the non-ideal transition bandof the low-pass filter in the microphone

We present UPS+ an enhanced UPS that can provide sub-centimeter indoor localization to current smart devices andoperate absolutely at ultrasonic spectrum which is consid-erably beyond the hearing system of humans and pets Ourkey innovation is the renewed efforts on promoting tradi-tional UPSs to become serviceable to ultrasound-incapablereceivers A recent finding shows that a combination oftwo ultrasounds at two frequencies (eg f1 and f2) mayget shifted to a lower differential frequency (eg | f1 minus f2 |)when they arrive simultaneously at a microphone [33ndash35]Toward this hardware property as a natural downconversionapproach we ldquopull downrdquo ultrasonic beacons to the audiblespectrum which the receiver can process To this end UPS+adopts a heterogeneous architecture that consists of two typesof custom-made beacon devices uBeacon and cBeacon as

shown in Fig 1 They broadcast beacons at two ultrasonicfrequencies thereby ensuring that the beacon signals willnot disturb humans or pets In particular a large number ofbattery-supplied uBeacons are deployed as location anchorswhereas a single cable-powered cBeacon is installed for theultrasonic downconversion Finally periodic beacons fromuBeacons are downconverted by the beacon signals fromthe cBeacon After capturing the beacons from at least fouruBeacons the receiver computes the time of arrival and thenlocates itself via trilateration

The nonlinearity effect has been verified and demonstratedin various acoustic attacks [33ndash35] However whether suchproperty can be used for indoor localization remains unclearunless the following three concerns are addressed

bull How can we deal with the frequency-selectivity Severalcopied ultrasounds with the same frequency but propagatedalong different paths may induce constructive or destructiveinterference at the receiver thereby leading to the frequencyselectivity issue The use of broadband acoustic signals (iechirps) is considered as an effective means to address thischallenge [24 29 36 37] However a uBeacon is made of anultrasonic transducer and only has a bandwidth of 2 kHzwhich is barely adequate to deal with the selectivity In thiswork we propose the technique of dynamical chirp spreadspectrum (DCCS) technique which spreads the single-tonepulses of the uBeacons over the cBeaconrsquos broadband chirps

bull How can the beacons advertised from multiple beacon de-vices be distinguished Trilateration requires the receiver toacquire beacons from at least four uBeacons thus multiplebeacon access is an unavoidable issue Frequency divisionmultiple address (FMDA) is typically adopted in previousUPSs However it fails in our scenario because each uBea-con sweeps an unpredictable dynamic band that dependson the receiverrsquos location which causes the downconvertedbeacons to overlap in the frequency-domain at the receiversInstead UPS+ is driven by a two-level multiple access mech-anism The receiver initially locates itself in a large spacethrough the chirp slope of the cBeacon and then decodes theuBeaconrsquos ID from its scheduled beacons

2

bullHow can the receiving energy of downconverted beacons beenhanced The nonlinearity effect is observed at the second-order harmonics which has an amplitude that is less thanthe fundamental signals Consequently the signal-to-noiseratio (SNR) of the beacon at the receiver side in UPS+ isless than that in a traditional UPS at the same position Wealleviate this problem through the following efforts On thetransmitting side a complicated custom-made driver circuitis designed to double the output power of the transducer Onthe receiving side a noise reduction algorithm is specially re-designed to turbocharge weak beacons by taking advantageof the dual-microphones in smart devices

Summary of Results Fig 2 shows the experiment plat-form and the prototype Our evaluation is performed on 8types of off-the-shelf smart devices (ie five smartphonesan iPad an iWatch and a pair of AirPods) The results demon-strate that UPS+ can fully utilize the 22 kHz bandwidth forchirps spread on these devices It performs localization withmedian and 90th percentile errors of 459 cm and 1457 cmrespectively Such accuracy matches or even exceeds someprevious UPSs The effective range (median error le 20 cm)of uBeacon equipped with three-transducers is around 6 mas equivalent to that of normal UPSs

Contribution This work presents UPS+ the first systemthat operates at an ultrasonic spectrum but remains service-able in currently available ultrasound-incapable smart de-vices The design of UPS+ introduces three key innovationsFirst it presents a centralized device called cBeacon to boostprevious UPSs Second it uses the nonlinearity effect of mi-crophone systems as an ultrasonic downconverter Third itspreads single-tone pulses with dynamic chirps in the airand turbocharges downconverted beacons at the receivingside The study also presents a prototype implementationand evaluation of UPS+ thereby demonstrating its accuracyin localizing smart devices in our office

2 RELATEDWORKThis work touches upon many topics related to indoor lo-calization nonlinearity effect and noise reduction Theseareas already have large bodies of research thus our reviewfocuses primarily on closely related works

(a) Sound-based Indoor Localization Sound-based in-door localization can be classified under broad categories ofrange-based [20 24 25 28ndash30 32 38ndash42] and range-free [1516 27 43 44 44ndash46] methods Here we focus only on range-based methods We refer to [47] for a comprehensive surveyon various indoor localization techniques This category ofsolutions computes distances based on how long sound takesto propagate between a sender and a receiver [20 24 25 28ndash30 32 38ndash41] For example Active Bats [38] and Cricket [20]

are pioneering range-based localization systems that use ul-trasonic beacons Dolphin [29] presents a new design forultrasonic transmitters and receivers BeepBeep [28] usesthe same range approach to estimate the distance betweentwo cellular phones [31] attempts to identify the location ofa mobile phone in a car using vehicle-mounted audio speak-ers ApneaApp [41] uses an estimated range to track humanbreath Pulse Compression(PC) [24] is the closest to our workbut it uses 19 sim 20 kHz band which may still be sensitivefor infants or pets By contrary our beacons are advertisedat 50 kHz or higher which results in zero noise pollution toindoor creatures Moreover our work uses dynamic chirps(sect5) to spread the spectrum instead of the static chirps usedin [24] In summary previous works either require ultrasonicreceivers or advertise audible beacons Additional compar-isons are provided in sect9(b) Nonlinearity Effect Nonlinearity has been explored

for many purposes [33ndash35] in recent years Backdoor [34]constructs an acoustic (but inaudible) communication chan-nel between two speakers and a microphone over ultrasoundbands with the nonlinearity effect DolphinAttack [33] andLipRead [35] utilize the nonlinearity effect to send inaudiblecommands to voice-enabled devices such as Amazon EchoOur system is inspired by these previous works Howeverwe use this effect for a different purpose and face challengesthat vary from those in previous studies

(c) Noise Reduction Finally several works [48ndash51] fromthe speech field havemotivated us to design a dual-microphoneenabled turbocharging algorithm In contrast to previousworks we use the features of beacon signals to determinetheir absence

3 TOWARDS NONLINEARITY AS ANULTRASONIC DOWNCONVERTER

In this section we provide the background of nonlinearityeffect and describe how UPS+ can take advantage of thisphenomenon for ultrasonic downconversion

31 Primer on Microphone SystemTodayrsquos smart devices adopt a generic pipeline for soundprocessing After being captured by the microphone soundis first magnified by amplifiers To capture audible soundsthe microphone system is designed to be sensitive to thespectrum of 0 sim 22 kHz by using a low-pass filter (LPF) toremove sounds that are higher than 22 kHz even when theyare recorded by the microphone Finally the sampling rateof the analog-to-digital converter is typically 441 kHz andthe digitized signalrsquos frequency is limited to below 22 kHzaccording to the Nyquist sampling theorem All moduleswithin the microphone system are supposed to be linearthus the output signals are linear combinations of the input

3

0 1 2 3 4 5Time(s)

9

10

11

12

13

Frequency(Khz)

(a) Pulse+CW

1 2 3 4 5 6Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

(b) CW+Chirps

0 1 2 3 4 5Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

Unsampled

(c) Pulse+Chirps

Fig 3 Downconverted beacons in the frequency-domain The built-in recorder of an iPhone 8 is used to capture ultrasonic beaconswith different settings (a) the uBeacon advertises pulses at 40 kHz whereas the cBeacon transmits a single-tone continuous wave (CW) at 50kHz (b) the uBeacon advertises a CW at 40 kHz while the cBeacon transmits continuous chirps from 45 sim 55 kHz and (c) the uBeaconadvertises pulses at 40 kHz whereas the cBeacon transmits chirps from 45 sim 55 kHz

However the sound amplifiers exhibit strong nonlinearityoutside the operating band In particular if f gt 25 kHz thenthe net recorded sound Sout can be expressed in terms of theinput sound S as follows

Sout(t)|f gt25 =infinsumi=1

GiS(t) = G1S(t) +G2S2(t) +G3S

3(t) + middot middot middot (1)

where Gi is the gain for the ith item The amplifier createsnew frequencies (ie harmonics) due to the nonlinearityproperty [52]

32 Nonlinearity EffectTo operate the microphone system in an ultrasonic bandwe use two off-the-shelf ultrasound speakers to simultane-ously play two sounds S1(t) = A1 cos(2π f1t) and S2(t) =A2 cos(2π f2t) at frequencies of f1 and f2 respectively wheref1 and f2 gt 22 kHz After magnification the output Sout canbe modeled as follows

S(t) =G1(S1 + S2) +G2(S1 + S2)2 +G2(S1 + S2)3 + middot middot middot (2)

We discuss only the second-order term by manipulating theinput signal S because the third- and higher-order termshave an extremely weak gain and can be disregarded Weexpand the second-order item as follows

G2(S1 + S2)2 =12(G2A

21 +G2A

22) minus

12(cos(2π f1t) + cos(2π f2t))

+G2A1A2 cos(2π (f1 + f2)t) minusG2A1A2 cos(2π (f1 minus f2)t) (3)

Interestingly four new frequencies (ie 2f1 2f2 f1 minus f2 andf1 + f2) are created after magnification All the ultrasonicfrequencies at f1 f2 2f1 2f2 and f1 + f2 are filtered out dueto the LPFrsquos cut off at 22 kHz However (f1 minus f2) remainsTranslating to actual numbers when f1 = 50 kHz and f2 = 40kHz the microphone will output acoustic signals at 10 kHzThe net effect is that a completely inaudible frequency isrecorded by unmodified off-the-shelf microphones therebyoffering natural downconversion for ultrasonic signals For

brevity we extract the downconverted sound signal denotedby Sdarr(t) as our target signals

Sdarr(t) = G2A1A2 cos(2π (f1 minus f2)t) = Adarr cos(2π (f1 minus f2)t) (4)

where Adarr = G2A1A2 The feasibility and accuracy of thenonlinearity effect on existing microphone systems havebeen completely verified and demonstrated in recent previ-ous works Here we omit to conduct our own verificationand refer readers to the results in [33ndash35] Especially [33]tests dozens of devices to show that the nonlinearity is aubiquitous effect across the currently available smart device

33 Ultrasonic DownconversionNonlinearity effect has been considered a type of ldquopollutionrdquoor a security ldquoback doorrdquo in previous works Instead weexplore this underlying physical property as a novel andpositive downconverter to engineer a practical UPS Suchdownconverter does not require receivers to be equippedwith any ultrasonic sensors It offers two clear advantagesFirst downconversion is generated by the hardware propertyof electronic amplifiers and thus it will not affect humansor animals Second the nonlinearity effect can downcon-vert ultrasonic beacons into any frequency below 22 kHzby manipulating two ultrasonic speakers thereby implicitlyproviding up to 22 kHz broadband Subsequently we willexplore this effect to design UPS+

4 SYSTEM DESIGNWe call the ultrasonic signal beacon and the devices that cantransmit ultrasonic beacons beacon device Our fundamen-tal concept is to deploy many beacon devices in the targetspace and utilize the nonlinearity effect to downconvert theirbeacons into an audible spectrum that can be recognized bysmart devices We present the system design at a high-levelin this section

4

Table 1 Comparisons of ultrasonic components

Type Size(cm2) Price BW Rng SupplyTransducer 16 times 16 1 $ 2 kHz 8 m Battery

Speaker 9 times 12 50 $ 200 kHz 50 m Cable

41 Dilemma in Engineering UPS+To take advantage of the nonlinearity the straightforwarddesign should equip each beacon device with two ultrasonicspeakers which transmit ultrasound at two different frequen-cies However this design is uneconomical and impracticalin engineering practices Our engineering philosophy is toutilize commercial off-the-shelf components to build the entiresystem and to benefit from the economies of scale by beingcost-effective In the present market two types of commercialultrasonic components namely transducers and speakers areavailable as listed in Table 1

bull Transducers Ultrasonic transducers (aka ultrasonic sen-sors) can convert AC into ultrasound or vice versa Theyare small low cost (1 $) and power-saving but suffer froma limited range (ie max at 8 m) and a narrow band (sim 2kHz) The majority of previous UPSs (eg Cricket) usetransducers to build beacon devices

bull Speaker Ultrasonic speakers are bulky and expensive (50$) but have a broadband of 200 kHz (ie 0 sim 200 kHz) anda propagation range of up to 50 m These speakers wereused as anchors in PC [24]

A comparison of their characteristics presents an engineeringdilemma in component selection that is deploying a largenumber of speakers is extremely expensive whereas cheaptransducers fail to satisfy our broadband demand

42 System ArchitectureTo overcome the dilemma we design a heterogeneous archi-tecture that jointly uses the two ultrasonic components byinventing two types of beacon devices

bull uBeacon A uBeacon is composed of an ultrasonic trans-ducer operating at a single tone (eg 40 kHz) with a di-rectivity of 30 The total cost of a uBeacon is about 5$We can deploy uBeacons at a large scale due to its lowcost uBeacons act as location anchors whose locationsare known ahead of time to receivers

bull cBeacon A cBeacon is composed of a powerful ultrasonicspeaker A single cBeacon is sufficient to cover a largetarget space (eg a lecture room) and shared by manyuBeacons since the speaker is wideband and long-rangePowered by electric cables a cBeacon produces continuousultrasonic signals and functions for downconversion Itcan be attached anywhere because its location is irrelevantto the localization results

cBeacon

uBeacon

Continuous Wave

Pulse

Audiblepulse

Pulse

tuBeacon

Fig 4 Downconversion timing

Such heterogeneous architecture achieves a trade-off be-tween the cost and effectiveness In particular we can im-mediately stop the potential interference to other recordingactivities (eg recording voice using smart phones) by turn-ing off the single cBeacon rather than stopping distributeduBeacons one by one or awaken cBeacon for localizationwhen needed

43 Timing for DownconversionWeassign two different ultrasonic frequencies to the uBeaconand the cBeacon and they are deployed in different locationsThe receiver can detect a downconverted beacon only whentheir ultrasonic beacons arrive simultaneously at the receiverA natural question is how UPS+ synchronizes the arrivalsof two types of beacons To do so the cBeacon persistentlytransmits a continuous wave (CW) whereas the uBeaconsadvertise short beacons every a few milliseconds Fig 4 usesa toy example to explain the design Given that signals fromthe cBeacon arrive continuously downconversion occursonly when a uBeaconrsquos signal arrives at the receiver In thismanner timing is dependent only on the signal arrival ofthe uBeacon and irrelevant of the cBeacon Furthermorewe demonstrate the timing through an actual experimentin Fig 3(a) where the uBeacon advertises pulses at 40 kHzwhereas the cBeacon transmits a single-tone CW at 50 kHzWe observe an apparent pattern of pulses at 10 Hz everysecond similar to the uBeaconrsquos advertising pattern at 40kHz although the cBeacon never stops its signals

44 Trilateration with ToA in UPS+Suppose that the ith uBeacon is located at position Ui andtransmits a beacon at time ts The receiver detects the arrivalof this beacon at time ti Thus the beacon takes timinusts secondsto propagate in air We can compute the pseudo-range asfollows

c times (ti minus ts ) = c times tb + |Ui minus P | (5)

where c is the speed of sound P(x y z) is the receiverrsquos loca-tion and tb is the clock difference between the receiver andthe uBeacons All the uBeacons are assumed to be well syn-chronized in time and advertise beacons at time ts c ti andUi are known to the receiver hence the preceding equation

5

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

cBeacon uBeacon

WiFi Chipset

Transducer (recv)

Transducer (trans)Beacon devices(a) (b) (c)

Fig 2 UPS+ experimental platform (a) Experimental scenario (b) two beacon devices cBeacon and uBeacon and (c) zoom-in image ofthe circuit board of the uBeacon

utilizes the ultrasonic sound as the ranging media UPS out-performs RF-based systems in terms of accuracy For exampleMIT launched Project Oxygen [19] in 2004 In this project apioneering UPS named Cricket [20] was developed to providecentimeter-level accuracy location service Subsequently alarge number of follow-up works [21ndash32] have further uti-lized sound or ultrasound for ranging or localization (seesect2) However despite the extensive efforts exerted on UPSresearch only a few of the resulting solutions have been ac-tually deployed in practice at present For example Cricketproject ceased updating 10 years ago One of the importantreasons why UPSs do not receive enough attention today isthat they suffer from a serious defect that is todayrsquos smartdevices lack ultrasonic sensors and therefore cannot receiveultrasonic beacons from UPSs [25 28 31] To bridge this gapmany follow-up works [25 28 31] have advised transmittingbeacons at the spectrum of 20 sim 22 kHz where 22 kHz isthe upper bound of currently available microphones Thesound within this band is supposedly inaudible to the major-ity of humans Nevertheless it is still harmful to infants andpets who are hyper-sensitive to the high-frequency soundsMoreover the signals near the upper bound may be seriouslydistorted and attenuated due to the non-ideal transition bandof the low-pass filter in the microphone

We present UPS+ an enhanced UPS that can provide sub-centimeter indoor localization to current smart devices andoperate absolutely at ultrasonic spectrum which is consid-erably beyond the hearing system of humans and pets Ourkey innovation is the renewed efforts on promoting tradi-tional UPSs to become serviceable to ultrasound-incapablereceivers A recent finding shows that a combination oftwo ultrasounds at two frequencies (eg f1 and f2) mayget shifted to a lower differential frequency (eg | f1 minus f2 |)when they arrive simultaneously at a microphone [33ndash35]Toward this hardware property as a natural downconversionapproach we ldquopull downrdquo ultrasonic beacons to the audiblespectrum which the receiver can process To this end UPS+adopts a heterogeneous architecture that consists of two typesof custom-made beacon devices uBeacon and cBeacon as

shown in Fig 1 They broadcast beacons at two ultrasonicfrequencies thereby ensuring that the beacon signals willnot disturb humans or pets In particular a large number ofbattery-supplied uBeacons are deployed as location anchorswhereas a single cable-powered cBeacon is installed for theultrasonic downconversion Finally periodic beacons fromuBeacons are downconverted by the beacon signals fromthe cBeacon After capturing the beacons from at least fouruBeacons the receiver computes the time of arrival and thenlocates itself via trilateration

The nonlinearity effect has been verified and demonstratedin various acoustic attacks [33ndash35] However whether suchproperty can be used for indoor localization remains unclearunless the following three concerns are addressed

bull How can we deal with the frequency-selectivity Severalcopied ultrasounds with the same frequency but propagatedalong different paths may induce constructive or destructiveinterference at the receiver thereby leading to the frequencyselectivity issue The use of broadband acoustic signals (iechirps) is considered as an effective means to address thischallenge [24 29 36 37] However a uBeacon is made of anultrasonic transducer and only has a bandwidth of 2 kHzwhich is barely adequate to deal with the selectivity In thiswork we propose the technique of dynamical chirp spreadspectrum (DCCS) technique which spreads the single-tonepulses of the uBeacons over the cBeaconrsquos broadband chirps

bull How can the beacons advertised from multiple beacon de-vices be distinguished Trilateration requires the receiver toacquire beacons from at least four uBeacons thus multiplebeacon access is an unavoidable issue Frequency divisionmultiple address (FMDA) is typically adopted in previousUPSs However it fails in our scenario because each uBea-con sweeps an unpredictable dynamic band that dependson the receiverrsquos location which causes the downconvertedbeacons to overlap in the frequency-domain at the receiversInstead UPS+ is driven by a two-level multiple access mech-anism The receiver initially locates itself in a large spacethrough the chirp slope of the cBeacon and then decodes theuBeaconrsquos ID from its scheduled beacons

2

bullHow can the receiving energy of downconverted beacons beenhanced The nonlinearity effect is observed at the second-order harmonics which has an amplitude that is less thanthe fundamental signals Consequently the signal-to-noiseratio (SNR) of the beacon at the receiver side in UPS+ isless than that in a traditional UPS at the same position Wealleviate this problem through the following efforts On thetransmitting side a complicated custom-made driver circuitis designed to double the output power of the transducer Onthe receiving side a noise reduction algorithm is specially re-designed to turbocharge weak beacons by taking advantageof the dual-microphones in smart devices

Summary of Results Fig 2 shows the experiment plat-form and the prototype Our evaluation is performed on 8types of off-the-shelf smart devices (ie five smartphonesan iPad an iWatch and a pair of AirPods) The results demon-strate that UPS+ can fully utilize the 22 kHz bandwidth forchirps spread on these devices It performs localization withmedian and 90th percentile errors of 459 cm and 1457 cmrespectively Such accuracy matches or even exceeds someprevious UPSs The effective range (median error le 20 cm)of uBeacon equipped with three-transducers is around 6 mas equivalent to that of normal UPSs

Contribution This work presents UPS+ the first systemthat operates at an ultrasonic spectrum but remains service-able in currently available ultrasound-incapable smart de-vices The design of UPS+ introduces three key innovationsFirst it presents a centralized device called cBeacon to boostprevious UPSs Second it uses the nonlinearity effect of mi-crophone systems as an ultrasonic downconverter Third itspreads single-tone pulses with dynamic chirps in the airand turbocharges downconverted beacons at the receivingside The study also presents a prototype implementationand evaluation of UPS+ thereby demonstrating its accuracyin localizing smart devices in our office

2 RELATEDWORKThis work touches upon many topics related to indoor lo-calization nonlinearity effect and noise reduction Theseareas already have large bodies of research thus our reviewfocuses primarily on closely related works

(a) Sound-based Indoor Localization Sound-based in-door localization can be classified under broad categories ofrange-based [20 24 25 28ndash30 32 38ndash42] and range-free [1516 27 43 44 44ndash46] methods Here we focus only on range-based methods We refer to [47] for a comprehensive surveyon various indoor localization techniques This category ofsolutions computes distances based on how long sound takesto propagate between a sender and a receiver [20 24 25 28ndash30 32 38ndash41] For example Active Bats [38] and Cricket [20]

are pioneering range-based localization systems that use ul-trasonic beacons Dolphin [29] presents a new design forultrasonic transmitters and receivers BeepBeep [28] usesthe same range approach to estimate the distance betweentwo cellular phones [31] attempts to identify the location ofa mobile phone in a car using vehicle-mounted audio speak-ers ApneaApp [41] uses an estimated range to track humanbreath Pulse Compression(PC) [24] is the closest to our workbut it uses 19 sim 20 kHz band which may still be sensitivefor infants or pets By contrary our beacons are advertisedat 50 kHz or higher which results in zero noise pollution toindoor creatures Moreover our work uses dynamic chirps(sect5) to spread the spectrum instead of the static chirps usedin [24] In summary previous works either require ultrasonicreceivers or advertise audible beacons Additional compar-isons are provided in sect9(b) Nonlinearity Effect Nonlinearity has been explored

for many purposes [33ndash35] in recent years Backdoor [34]constructs an acoustic (but inaudible) communication chan-nel between two speakers and a microphone over ultrasoundbands with the nonlinearity effect DolphinAttack [33] andLipRead [35] utilize the nonlinearity effect to send inaudiblecommands to voice-enabled devices such as Amazon EchoOur system is inspired by these previous works Howeverwe use this effect for a different purpose and face challengesthat vary from those in previous studies

(c) Noise Reduction Finally several works [48ndash51] fromthe speech field havemotivated us to design a dual-microphoneenabled turbocharging algorithm In contrast to previousworks we use the features of beacon signals to determinetheir absence

3 TOWARDS NONLINEARITY AS ANULTRASONIC DOWNCONVERTER

In this section we provide the background of nonlinearityeffect and describe how UPS+ can take advantage of thisphenomenon for ultrasonic downconversion

31 Primer on Microphone SystemTodayrsquos smart devices adopt a generic pipeline for soundprocessing After being captured by the microphone soundis first magnified by amplifiers To capture audible soundsthe microphone system is designed to be sensitive to thespectrum of 0 sim 22 kHz by using a low-pass filter (LPF) toremove sounds that are higher than 22 kHz even when theyare recorded by the microphone Finally the sampling rateof the analog-to-digital converter is typically 441 kHz andthe digitized signalrsquos frequency is limited to below 22 kHzaccording to the Nyquist sampling theorem All moduleswithin the microphone system are supposed to be linearthus the output signals are linear combinations of the input

3

0 1 2 3 4 5Time(s)

9

10

11

12

13

Frequency(Khz)

(a) Pulse+CW

1 2 3 4 5 6Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

(b) CW+Chirps

0 1 2 3 4 5Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

Unsampled

(c) Pulse+Chirps

Fig 3 Downconverted beacons in the frequency-domain The built-in recorder of an iPhone 8 is used to capture ultrasonic beaconswith different settings (a) the uBeacon advertises pulses at 40 kHz whereas the cBeacon transmits a single-tone continuous wave (CW) at 50kHz (b) the uBeacon advertises a CW at 40 kHz while the cBeacon transmits continuous chirps from 45 sim 55 kHz and (c) the uBeaconadvertises pulses at 40 kHz whereas the cBeacon transmits chirps from 45 sim 55 kHz

However the sound amplifiers exhibit strong nonlinearityoutside the operating band In particular if f gt 25 kHz thenthe net recorded sound Sout can be expressed in terms of theinput sound S as follows

Sout(t)|f gt25 =infinsumi=1

GiS(t) = G1S(t) +G2S2(t) +G3S

3(t) + middot middot middot (1)

where Gi is the gain for the ith item The amplifier createsnew frequencies (ie harmonics) due to the nonlinearityproperty [52]

32 Nonlinearity EffectTo operate the microphone system in an ultrasonic bandwe use two off-the-shelf ultrasound speakers to simultane-ously play two sounds S1(t) = A1 cos(2π f1t) and S2(t) =A2 cos(2π f2t) at frequencies of f1 and f2 respectively wheref1 and f2 gt 22 kHz After magnification the output Sout canbe modeled as follows

S(t) =G1(S1 + S2) +G2(S1 + S2)2 +G2(S1 + S2)3 + middot middot middot (2)

We discuss only the second-order term by manipulating theinput signal S because the third- and higher-order termshave an extremely weak gain and can be disregarded Weexpand the second-order item as follows

G2(S1 + S2)2 =12(G2A

21 +G2A

22) minus

12(cos(2π f1t) + cos(2π f2t))

+G2A1A2 cos(2π (f1 + f2)t) minusG2A1A2 cos(2π (f1 minus f2)t) (3)

Interestingly four new frequencies (ie 2f1 2f2 f1 minus f2 andf1 + f2) are created after magnification All the ultrasonicfrequencies at f1 f2 2f1 2f2 and f1 + f2 are filtered out dueto the LPFrsquos cut off at 22 kHz However (f1 minus f2) remainsTranslating to actual numbers when f1 = 50 kHz and f2 = 40kHz the microphone will output acoustic signals at 10 kHzThe net effect is that a completely inaudible frequency isrecorded by unmodified off-the-shelf microphones therebyoffering natural downconversion for ultrasonic signals For

brevity we extract the downconverted sound signal denotedby Sdarr(t) as our target signals

Sdarr(t) = G2A1A2 cos(2π (f1 minus f2)t) = Adarr cos(2π (f1 minus f2)t) (4)

where Adarr = G2A1A2 The feasibility and accuracy of thenonlinearity effect on existing microphone systems havebeen completely verified and demonstrated in recent previ-ous works Here we omit to conduct our own verificationand refer readers to the results in [33ndash35] Especially [33]tests dozens of devices to show that the nonlinearity is aubiquitous effect across the currently available smart device

33 Ultrasonic DownconversionNonlinearity effect has been considered a type of ldquopollutionrdquoor a security ldquoback doorrdquo in previous works Instead weexplore this underlying physical property as a novel andpositive downconverter to engineer a practical UPS Suchdownconverter does not require receivers to be equippedwith any ultrasonic sensors It offers two clear advantagesFirst downconversion is generated by the hardware propertyof electronic amplifiers and thus it will not affect humansor animals Second the nonlinearity effect can downcon-vert ultrasonic beacons into any frequency below 22 kHzby manipulating two ultrasonic speakers thereby implicitlyproviding up to 22 kHz broadband Subsequently we willexplore this effect to design UPS+

4 SYSTEM DESIGNWe call the ultrasonic signal beacon and the devices that cantransmit ultrasonic beacons beacon device Our fundamen-tal concept is to deploy many beacon devices in the targetspace and utilize the nonlinearity effect to downconvert theirbeacons into an audible spectrum that can be recognized bysmart devices We present the system design at a high-levelin this section

4

Table 1 Comparisons of ultrasonic components

Type Size(cm2) Price BW Rng SupplyTransducer 16 times 16 1 $ 2 kHz 8 m Battery

Speaker 9 times 12 50 $ 200 kHz 50 m Cable

41 Dilemma in Engineering UPS+To take advantage of the nonlinearity the straightforwarddesign should equip each beacon device with two ultrasonicspeakers which transmit ultrasound at two different frequen-cies However this design is uneconomical and impracticalin engineering practices Our engineering philosophy is toutilize commercial off-the-shelf components to build the entiresystem and to benefit from the economies of scale by beingcost-effective In the present market two types of commercialultrasonic components namely transducers and speakers areavailable as listed in Table 1

bull Transducers Ultrasonic transducers (aka ultrasonic sen-sors) can convert AC into ultrasound or vice versa Theyare small low cost (1 $) and power-saving but suffer froma limited range (ie max at 8 m) and a narrow band (sim 2kHz) The majority of previous UPSs (eg Cricket) usetransducers to build beacon devices

bull Speaker Ultrasonic speakers are bulky and expensive (50$) but have a broadband of 200 kHz (ie 0 sim 200 kHz) anda propagation range of up to 50 m These speakers wereused as anchors in PC [24]

A comparison of their characteristics presents an engineeringdilemma in component selection that is deploying a largenumber of speakers is extremely expensive whereas cheaptransducers fail to satisfy our broadband demand

42 System ArchitectureTo overcome the dilemma we design a heterogeneous archi-tecture that jointly uses the two ultrasonic components byinventing two types of beacon devices

bull uBeacon A uBeacon is composed of an ultrasonic trans-ducer operating at a single tone (eg 40 kHz) with a di-rectivity of 30 The total cost of a uBeacon is about 5$We can deploy uBeacons at a large scale due to its lowcost uBeacons act as location anchors whose locationsare known ahead of time to receivers

bull cBeacon A cBeacon is composed of a powerful ultrasonicspeaker A single cBeacon is sufficient to cover a largetarget space (eg a lecture room) and shared by manyuBeacons since the speaker is wideband and long-rangePowered by electric cables a cBeacon produces continuousultrasonic signals and functions for downconversion Itcan be attached anywhere because its location is irrelevantto the localization results

cBeacon

uBeacon

Continuous Wave

Pulse

Audiblepulse

Pulse

tuBeacon

Fig 4 Downconversion timing

Such heterogeneous architecture achieves a trade-off be-tween the cost and effectiveness In particular we can im-mediately stop the potential interference to other recordingactivities (eg recording voice using smart phones) by turn-ing off the single cBeacon rather than stopping distributeduBeacons one by one or awaken cBeacon for localizationwhen needed

43 Timing for DownconversionWeassign two different ultrasonic frequencies to the uBeaconand the cBeacon and they are deployed in different locationsThe receiver can detect a downconverted beacon only whentheir ultrasonic beacons arrive simultaneously at the receiverA natural question is how UPS+ synchronizes the arrivalsof two types of beacons To do so the cBeacon persistentlytransmits a continuous wave (CW) whereas the uBeaconsadvertise short beacons every a few milliseconds Fig 4 usesa toy example to explain the design Given that signals fromthe cBeacon arrive continuously downconversion occursonly when a uBeaconrsquos signal arrives at the receiver In thismanner timing is dependent only on the signal arrival ofthe uBeacon and irrelevant of the cBeacon Furthermorewe demonstrate the timing through an actual experimentin Fig 3(a) where the uBeacon advertises pulses at 40 kHzwhereas the cBeacon transmits a single-tone CW at 50 kHzWe observe an apparent pattern of pulses at 10 Hz everysecond similar to the uBeaconrsquos advertising pattern at 40kHz although the cBeacon never stops its signals

44 Trilateration with ToA in UPS+Suppose that the ith uBeacon is located at position Ui andtransmits a beacon at time ts The receiver detects the arrivalof this beacon at time ti Thus the beacon takes timinusts secondsto propagate in air We can compute the pseudo-range asfollows

c times (ti minus ts ) = c times tb + |Ui minus P | (5)

where c is the speed of sound P(x y z) is the receiverrsquos loca-tion and tb is the clock difference between the receiver andthe uBeacons All the uBeacons are assumed to be well syn-chronized in time and advertise beacons at time ts c ti andUi are known to the receiver hence the preceding equation

5

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

bullHow can the receiving energy of downconverted beacons beenhanced The nonlinearity effect is observed at the second-order harmonics which has an amplitude that is less thanthe fundamental signals Consequently the signal-to-noiseratio (SNR) of the beacon at the receiver side in UPS+ isless than that in a traditional UPS at the same position Wealleviate this problem through the following efforts On thetransmitting side a complicated custom-made driver circuitis designed to double the output power of the transducer Onthe receiving side a noise reduction algorithm is specially re-designed to turbocharge weak beacons by taking advantageof the dual-microphones in smart devices

Summary of Results Fig 2 shows the experiment plat-form and the prototype Our evaluation is performed on 8types of off-the-shelf smart devices (ie five smartphonesan iPad an iWatch and a pair of AirPods) The results demon-strate that UPS+ can fully utilize the 22 kHz bandwidth forchirps spread on these devices It performs localization withmedian and 90th percentile errors of 459 cm and 1457 cmrespectively Such accuracy matches or even exceeds someprevious UPSs The effective range (median error le 20 cm)of uBeacon equipped with three-transducers is around 6 mas equivalent to that of normal UPSs

Contribution This work presents UPS+ the first systemthat operates at an ultrasonic spectrum but remains service-able in currently available ultrasound-incapable smart de-vices The design of UPS+ introduces three key innovationsFirst it presents a centralized device called cBeacon to boostprevious UPSs Second it uses the nonlinearity effect of mi-crophone systems as an ultrasonic downconverter Third itspreads single-tone pulses with dynamic chirps in the airand turbocharges downconverted beacons at the receivingside The study also presents a prototype implementationand evaluation of UPS+ thereby demonstrating its accuracyin localizing smart devices in our office

2 RELATEDWORKThis work touches upon many topics related to indoor lo-calization nonlinearity effect and noise reduction Theseareas already have large bodies of research thus our reviewfocuses primarily on closely related works

(a) Sound-based Indoor Localization Sound-based in-door localization can be classified under broad categories ofrange-based [20 24 25 28ndash30 32 38ndash42] and range-free [1516 27 43 44 44ndash46] methods Here we focus only on range-based methods We refer to [47] for a comprehensive surveyon various indoor localization techniques This category ofsolutions computes distances based on how long sound takesto propagate between a sender and a receiver [20 24 25 28ndash30 32 38ndash41] For example Active Bats [38] and Cricket [20]

are pioneering range-based localization systems that use ul-trasonic beacons Dolphin [29] presents a new design forultrasonic transmitters and receivers BeepBeep [28] usesthe same range approach to estimate the distance betweentwo cellular phones [31] attempts to identify the location ofa mobile phone in a car using vehicle-mounted audio speak-ers ApneaApp [41] uses an estimated range to track humanbreath Pulse Compression(PC) [24] is the closest to our workbut it uses 19 sim 20 kHz band which may still be sensitivefor infants or pets By contrary our beacons are advertisedat 50 kHz or higher which results in zero noise pollution toindoor creatures Moreover our work uses dynamic chirps(sect5) to spread the spectrum instead of the static chirps usedin [24] In summary previous works either require ultrasonicreceivers or advertise audible beacons Additional compar-isons are provided in sect9(b) Nonlinearity Effect Nonlinearity has been explored

for many purposes [33ndash35] in recent years Backdoor [34]constructs an acoustic (but inaudible) communication chan-nel between two speakers and a microphone over ultrasoundbands with the nonlinearity effect DolphinAttack [33] andLipRead [35] utilize the nonlinearity effect to send inaudiblecommands to voice-enabled devices such as Amazon EchoOur system is inspired by these previous works Howeverwe use this effect for a different purpose and face challengesthat vary from those in previous studies

(c) Noise Reduction Finally several works [48ndash51] fromthe speech field havemotivated us to design a dual-microphoneenabled turbocharging algorithm In contrast to previousworks we use the features of beacon signals to determinetheir absence

3 TOWARDS NONLINEARITY AS ANULTRASONIC DOWNCONVERTER

In this section we provide the background of nonlinearityeffect and describe how UPS+ can take advantage of thisphenomenon for ultrasonic downconversion

31 Primer on Microphone SystemTodayrsquos smart devices adopt a generic pipeline for soundprocessing After being captured by the microphone soundis first magnified by amplifiers To capture audible soundsthe microphone system is designed to be sensitive to thespectrum of 0 sim 22 kHz by using a low-pass filter (LPF) toremove sounds that are higher than 22 kHz even when theyare recorded by the microphone Finally the sampling rateof the analog-to-digital converter is typically 441 kHz andthe digitized signalrsquos frequency is limited to below 22 kHzaccording to the Nyquist sampling theorem All moduleswithin the microphone system are supposed to be linearthus the output signals are linear combinations of the input

3

0 1 2 3 4 5Time(s)

9

10

11

12

13

Frequency(Khz)

(a) Pulse+CW

1 2 3 4 5 6Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

(b) CW+Chirps

0 1 2 3 4 5Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

Unsampled

(c) Pulse+Chirps

Fig 3 Downconverted beacons in the frequency-domain The built-in recorder of an iPhone 8 is used to capture ultrasonic beaconswith different settings (a) the uBeacon advertises pulses at 40 kHz whereas the cBeacon transmits a single-tone continuous wave (CW) at 50kHz (b) the uBeacon advertises a CW at 40 kHz while the cBeacon transmits continuous chirps from 45 sim 55 kHz and (c) the uBeaconadvertises pulses at 40 kHz whereas the cBeacon transmits chirps from 45 sim 55 kHz

However the sound amplifiers exhibit strong nonlinearityoutside the operating band In particular if f gt 25 kHz thenthe net recorded sound Sout can be expressed in terms of theinput sound S as follows

Sout(t)|f gt25 =infinsumi=1

GiS(t) = G1S(t) +G2S2(t) +G3S

3(t) + middot middot middot (1)

where Gi is the gain for the ith item The amplifier createsnew frequencies (ie harmonics) due to the nonlinearityproperty [52]

32 Nonlinearity EffectTo operate the microphone system in an ultrasonic bandwe use two off-the-shelf ultrasound speakers to simultane-ously play two sounds S1(t) = A1 cos(2π f1t) and S2(t) =A2 cos(2π f2t) at frequencies of f1 and f2 respectively wheref1 and f2 gt 22 kHz After magnification the output Sout canbe modeled as follows

S(t) =G1(S1 + S2) +G2(S1 + S2)2 +G2(S1 + S2)3 + middot middot middot (2)

We discuss only the second-order term by manipulating theinput signal S because the third- and higher-order termshave an extremely weak gain and can be disregarded Weexpand the second-order item as follows

G2(S1 + S2)2 =12(G2A

21 +G2A

22) minus

12(cos(2π f1t) + cos(2π f2t))

+G2A1A2 cos(2π (f1 + f2)t) minusG2A1A2 cos(2π (f1 minus f2)t) (3)

Interestingly four new frequencies (ie 2f1 2f2 f1 minus f2 andf1 + f2) are created after magnification All the ultrasonicfrequencies at f1 f2 2f1 2f2 and f1 + f2 are filtered out dueto the LPFrsquos cut off at 22 kHz However (f1 minus f2) remainsTranslating to actual numbers when f1 = 50 kHz and f2 = 40kHz the microphone will output acoustic signals at 10 kHzThe net effect is that a completely inaudible frequency isrecorded by unmodified off-the-shelf microphones therebyoffering natural downconversion for ultrasonic signals For

brevity we extract the downconverted sound signal denotedby Sdarr(t) as our target signals

Sdarr(t) = G2A1A2 cos(2π (f1 minus f2)t) = Adarr cos(2π (f1 minus f2)t) (4)

where Adarr = G2A1A2 The feasibility and accuracy of thenonlinearity effect on existing microphone systems havebeen completely verified and demonstrated in recent previ-ous works Here we omit to conduct our own verificationand refer readers to the results in [33ndash35] Especially [33]tests dozens of devices to show that the nonlinearity is aubiquitous effect across the currently available smart device

33 Ultrasonic DownconversionNonlinearity effect has been considered a type of ldquopollutionrdquoor a security ldquoback doorrdquo in previous works Instead weexplore this underlying physical property as a novel andpositive downconverter to engineer a practical UPS Suchdownconverter does not require receivers to be equippedwith any ultrasonic sensors It offers two clear advantagesFirst downconversion is generated by the hardware propertyof electronic amplifiers and thus it will not affect humansor animals Second the nonlinearity effect can downcon-vert ultrasonic beacons into any frequency below 22 kHzby manipulating two ultrasonic speakers thereby implicitlyproviding up to 22 kHz broadband Subsequently we willexplore this effect to design UPS+

4 SYSTEM DESIGNWe call the ultrasonic signal beacon and the devices that cantransmit ultrasonic beacons beacon device Our fundamen-tal concept is to deploy many beacon devices in the targetspace and utilize the nonlinearity effect to downconvert theirbeacons into an audible spectrum that can be recognized bysmart devices We present the system design at a high-levelin this section

4

Table 1 Comparisons of ultrasonic components

Type Size(cm2) Price BW Rng SupplyTransducer 16 times 16 1 $ 2 kHz 8 m Battery

Speaker 9 times 12 50 $ 200 kHz 50 m Cable

41 Dilemma in Engineering UPS+To take advantage of the nonlinearity the straightforwarddesign should equip each beacon device with two ultrasonicspeakers which transmit ultrasound at two different frequen-cies However this design is uneconomical and impracticalin engineering practices Our engineering philosophy is toutilize commercial off-the-shelf components to build the entiresystem and to benefit from the economies of scale by beingcost-effective In the present market two types of commercialultrasonic components namely transducers and speakers areavailable as listed in Table 1

bull Transducers Ultrasonic transducers (aka ultrasonic sen-sors) can convert AC into ultrasound or vice versa Theyare small low cost (1 $) and power-saving but suffer froma limited range (ie max at 8 m) and a narrow band (sim 2kHz) The majority of previous UPSs (eg Cricket) usetransducers to build beacon devices

bull Speaker Ultrasonic speakers are bulky and expensive (50$) but have a broadband of 200 kHz (ie 0 sim 200 kHz) anda propagation range of up to 50 m These speakers wereused as anchors in PC [24]

A comparison of their characteristics presents an engineeringdilemma in component selection that is deploying a largenumber of speakers is extremely expensive whereas cheaptransducers fail to satisfy our broadband demand

42 System ArchitectureTo overcome the dilemma we design a heterogeneous archi-tecture that jointly uses the two ultrasonic components byinventing two types of beacon devices

bull uBeacon A uBeacon is composed of an ultrasonic trans-ducer operating at a single tone (eg 40 kHz) with a di-rectivity of 30 The total cost of a uBeacon is about 5$We can deploy uBeacons at a large scale due to its lowcost uBeacons act as location anchors whose locationsare known ahead of time to receivers

bull cBeacon A cBeacon is composed of a powerful ultrasonicspeaker A single cBeacon is sufficient to cover a largetarget space (eg a lecture room) and shared by manyuBeacons since the speaker is wideband and long-rangePowered by electric cables a cBeacon produces continuousultrasonic signals and functions for downconversion Itcan be attached anywhere because its location is irrelevantto the localization results

cBeacon

uBeacon

Continuous Wave

Pulse

Audiblepulse

Pulse

tuBeacon

Fig 4 Downconversion timing

Such heterogeneous architecture achieves a trade-off be-tween the cost and effectiveness In particular we can im-mediately stop the potential interference to other recordingactivities (eg recording voice using smart phones) by turn-ing off the single cBeacon rather than stopping distributeduBeacons one by one or awaken cBeacon for localizationwhen needed

43 Timing for DownconversionWeassign two different ultrasonic frequencies to the uBeaconand the cBeacon and they are deployed in different locationsThe receiver can detect a downconverted beacon only whentheir ultrasonic beacons arrive simultaneously at the receiverA natural question is how UPS+ synchronizes the arrivalsof two types of beacons To do so the cBeacon persistentlytransmits a continuous wave (CW) whereas the uBeaconsadvertise short beacons every a few milliseconds Fig 4 usesa toy example to explain the design Given that signals fromthe cBeacon arrive continuously downconversion occursonly when a uBeaconrsquos signal arrives at the receiver In thismanner timing is dependent only on the signal arrival ofthe uBeacon and irrelevant of the cBeacon Furthermorewe demonstrate the timing through an actual experimentin Fig 3(a) where the uBeacon advertises pulses at 40 kHzwhereas the cBeacon transmits a single-tone CW at 50 kHzWe observe an apparent pattern of pulses at 10 Hz everysecond similar to the uBeaconrsquos advertising pattern at 40kHz although the cBeacon never stops its signals

44 Trilateration with ToA in UPS+Suppose that the ith uBeacon is located at position Ui andtransmits a beacon at time ts The receiver detects the arrivalof this beacon at time ti Thus the beacon takes timinusts secondsto propagate in air We can compute the pseudo-range asfollows

c times (ti minus ts ) = c times tb + |Ui minus P | (5)

where c is the speed of sound P(x y z) is the receiverrsquos loca-tion and tb is the clock difference between the receiver andthe uBeacons All the uBeacons are assumed to be well syn-chronized in time and advertise beacons at time ts c ti andUi are known to the receiver hence the preceding equation

5

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

0 1 2 3 4 5Time(s)

9

10

11

12

13

Frequency(Khz)

(a) Pulse+CW

1 2 3 4 5 6Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

(b) CW+Chirps

0 1 2 3 4 5Time(s)

5

7

9

11

13

15

17

Frequency(Khz)

Unsampled

(c) Pulse+Chirps

Fig 3 Downconverted beacons in the frequency-domain The built-in recorder of an iPhone 8 is used to capture ultrasonic beaconswith different settings (a) the uBeacon advertises pulses at 40 kHz whereas the cBeacon transmits a single-tone continuous wave (CW) at 50kHz (b) the uBeacon advertises a CW at 40 kHz while the cBeacon transmits continuous chirps from 45 sim 55 kHz and (c) the uBeaconadvertises pulses at 40 kHz whereas the cBeacon transmits chirps from 45 sim 55 kHz

However the sound amplifiers exhibit strong nonlinearityoutside the operating band In particular if f gt 25 kHz thenthe net recorded sound Sout can be expressed in terms of theinput sound S as follows

Sout(t)|f gt25 =infinsumi=1

GiS(t) = G1S(t) +G2S2(t) +G3S

3(t) + middot middot middot (1)

where Gi is the gain for the ith item The amplifier createsnew frequencies (ie harmonics) due to the nonlinearityproperty [52]

32 Nonlinearity EffectTo operate the microphone system in an ultrasonic bandwe use two off-the-shelf ultrasound speakers to simultane-ously play two sounds S1(t) = A1 cos(2π f1t) and S2(t) =A2 cos(2π f2t) at frequencies of f1 and f2 respectively wheref1 and f2 gt 22 kHz After magnification the output Sout canbe modeled as follows

S(t) =G1(S1 + S2) +G2(S1 + S2)2 +G2(S1 + S2)3 + middot middot middot (2)

We discuss only the second-order term by manipulating theinput signal S because the third- and higher-order termshave an extremely weak gain and can be disregarded Weexpand the second-order item as follows

G2(S1 + S2)2 =12(G2A

21 +G2A

22) minus

12(cos(2π f1t) + cos(2π f2t))

+G2A1A2 cos(2π (f1 + f2)t) minusG2A1A2 cos(2π (f1 minus f2)t) (3)

Interestingly four new frequencies (ie 2f1 2f2 f1 minus f2 andf1 + f2) are created after magnification All the ultrasonicfrequencies at f1 f2 2f1 2f2 and f1 + f2 are filtered out dueto the LPFrsquos cut off at 22 kHz However (f1 minus f2) remainsTranslating to actual numbers when f1 = 50 kHz and f2 = 40kHz the microphone will output acoustic signals at 10 kHzThe net effect is that a completely inaudible frequency isrecorded by unmodified off-the-shelf microphones therebyoffering natural downconversion for ultrasonic signals For

brevity we extract the downconverted sound signal denotedby Sdarr(t) as our target signals

Sdarr(t) = G2A1A2 cos(2π (f1 minus f2)t) = Adarr cos(2π (f1 minus f2)t) (4)

where Adarr = G2A1A2 The feasibility and accuracy of thenonlinearity effect on existing microphone systems havebeen completely verified and demonstrated in recent previ-ous works Here we omit to conduct our own verificationand refer readers to the results in [33ndash35] Especially [33]tests dozens of devices to show that the nonlinearity is aubiquitous effect across the currently available smart device

33 Ultrasonic DownconversionNonlinearity effect has been considered a type of ldquopollutionrdquoor a security ldquoback doorrdquo in previous works Instead weexplore this underlying physical property as a novel andpositive downconverter to engineer a practical UPS Suchdownconverter does not require receivers to be equippedwith any ultrasonic sensors It offers two clear advantagesFirst downconversion is generated by the hardware propertyof electronic amplifiers and thus it will not affect humansor animals Second the nonlinearity effect can downcon-vert ultrasonic beacons into any frequency below 22 kHzby manipulating two ultrasonic speakers thereby implicitlyproviding up to 22 kHz broadband Subsequently we willexplore this effect to design UPS+

4 SYSTEM DESIGNWe call the ultrasonic signal beacon and the devices that cantransmit ultrasonic beacons beacon device Our fundamen-tal concept is to deploy many beacon devices in the targetspace and utilize the nonlinearity effect to downconvert theirbeacons into an audible spectrum that can be recognized bysmart devices We present the system design at a high-levelin this section

4

Table 1 Comparisons of ultrasonic components

Type Size(cm2) Price BW Rng SupplyTransducer 16 times 16 1 $ 2 kHz 8 m Battery

Speaker 9 times 12 50 $ 200 kHz 50 m Cable

41 Dilemma in Engineering UPS+To take advantage of the nonlinearity the straightforwarddesign should equip each beacon device with two ultrasonicspeakers which transmit ultrasound at two different frequen-cies However this design is uneconomical and impracticalin engineering practices Our engineering philosophy is toutilize commercial off-the-shelf components to build the entiresystem and to benefit from the economies of scale by beingcost-effective In the present market two types of commercialultrasonic components namely transducers and speakers areavailable as listed in Table 1

bull Transducers Ultrasonic transducers (aka ultrasonic sen-sors) can convert AC into ultrasound or vice versa Theyare small low cost (1 $) and power-saving but suffer froma limited range (ie max at 8 m) and a narrow band (sim 2kHz) The majority of previous UPSs (eg Cricket) usetransducers to build beacon devices

bull Speaker Ultrasonic speakers are bulky and expensive (50$) but have a broadband of 200 kHz (ie 0 sim 200 kHz) anda propagation range of up to 50 m These speakers wereused as anchors in PC [24]

A comparison of their characteristics presents an engineeringdilemma in component selection that is deploying a largenumber of speakers is extremely expensive whereas cheaptransducers fail to satisfy our broadband demand

42 System ArchitectureTo overcome the dilemma we design a heterogeneous archi-tecture that jointly uses the two ultrasonic components byinventing two types of beacon devices

bull uBeacon A uBeacon is composed of an ultrasonic trans-ducer operating at a single tone (eg 40 kHz) with a di-rectivity of 30 The total cost of a uBeacon is about 5$We can deploy uBeacons at a large scale due to its lowcost uBeacons act as location anchors whose locationsare known ahead of time to receivers

bull cBeacon A cBeacon is composed of a powerful ultrasonicspeaker A single cBeacon is sufficient to cover a largetarget space (eg a lecture room) and shared by manyuBeacons since the speaker is wideband and long-rangePowered by electric cables a cBeacon produces continuousultrasonic signals and functions for downconversion Itcan be attached anywhere because its location is irrelevantto the localization results

cBeacon

uBeacon

Continuous Wave

Pulse

Audiblepulse

Pulse

tuBeacon

Fig 4 Downconversion timing

Such heterogeneous architecture achieves a trade-off be-tween the cost and effectiveness In particular we can im-mediately stop the potential interference to other recordingactivities (eg recording voice using smart phones) by turn-ing off the single cBeacon rather than stopping distributeduBeacons one by one or awaken cBeacon for localizationwhen needed

43 Timing for DownconversionWeassign two different ultrasonic frequencies to the uBeaconand the cBeacon and they are deployed in different locationsThe receiver can detect a downconverted beacon only whentheir ultrasonic beacons arrive simultaneously at the receiverA natural question is how UPS+ synchronizes the arrivalsof two types of beacons To do so the cBeacon persistentlytransmits a continuous wave (CW) whereas the uBeaconsadvertise short beacons every a few milliseconds Fig 4 usesa toy example to explain the design Given that signals fromthe cBeacon arrive continuously downconversion occursonly when a uBeaconrsquos signal arrives at the receiver In thismanner timing is dependent only on the signal arrival ofthe uBeacon and irrelevant of the cBeacon Furthermorewe demonstrate the timing through an actual experimentin Fig 3(a) where the uBeacon advertises pulses at 40 kHzwhereas the cBeacon transmits a single-tone CW at 50 kHzWe observe an apparent pattern of pulses at 10 Hz everysecond similar to the uBeaconrsquos advertising pattern at 40kHz although the cBeacon never stops its signals

44 Trilateration with ToA in UPS+Suppose that the ith uBeacon is located at position Ui andtransmits a beacon at time ts The receiver detects the arrivalof this beacon at time ti Thus the beacon takes timinusts secondsto propagate in air We can compute the pseudo-range asfollows

c times (ti minus ts ) = c times tb + |Ui minus P | (5)

where c is the speed of sound P(x y z) is the receiverrsquos loca-tion and tb is the clock difference between the receiver andthe uBeacons All the uBeacons are assumed to be well syn-chronized in time and advertise beacons at time ts c ti andUi are known to the receiver hence the preceding equation

5

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

Table 1 Comparisons of ultrasonic components

Type Size(cm2) Price BW Rng SupplyTransducer 16 times 16 1 $ 2 kHz 8 m Battery

Speaker 9 times 12 50 $ 200 kHz 50 m Cable

41 Dilemma in Engineering UPS+To take advantage of the nonlinearity the straightforwarddesign should equip each beacon device with two ultrasonicspeakers which transmit ultrasound at two different frequen-cies However this design is uneconomical and impracticalin engineering practices Our engineering philosophy is toutilize commercial off-the-shelf components to build the entiresystem and to benefit from the economies of scale by beingcost-effective In the present market two types of commercialultrasonic components namely transducers and speakers areavailable as listed in Table 1

bull Transducers Ultrasonic transducers (aka ultrasonic sen-sors) can convert AC into ultrasound or vice versa Theyare small low cost (1 $) and power-saving but suffer froma limited range (ie max at 8 m) and a narrow band (sim 2kHz) The majority of previous UPSs (eg Cricket) usetransducers to build beacon devices

bull Speaker Ultrasonic speakers are bulky and expensive (50$) but have a broadband of 200 kHz (ie 0 sim 200 kHz) anda propagation range of up to 50 m These speakers wereused as anchors in PC [24]

A comparison of their characteristics presents an engineeringdilemma in component selection that is deploying a largenumber of speakers is extremely expensive whereas cheaptransducers fail to satisfy our broadband demand

42 System ArchitectureTo overcome the dilemma we design a heterogeneous archi-tecture that jointly uses the two ultrasonic components byinventing two types of beacon devices

bull uBeacon A uBeacon is composed of an ultrasonic trans-ducer operating at a single tone (eg 40 kHz) with a di-rectivity of 30 The total cost of a uBeacon is about 5$We can deploy uBeacons at a large scale due to its lowcost uBeacons act as location anchors whose locationsare known ahead of time to receivers

bull cBeacon A cBeacon is composed of a powerful ultrasonicspeaker A single cBeacon is sufficient to cover a largetarget space (eg a lecture room) and shared by manyuBeacons since the speaker is wideband and long-rangePowered by electric cables a cBeacon produces continuousultrasonic signals and functions for downconversion Itcan be attached anywhere because its location is irrelevantto the localization results

cBeacon

uBeacon

Continuous Wave

Pulse

Audiblepulse

Pulse

tuBeacon

Fig 4 Downconversion timing

Such heterogeneous architecture achieves a trade-off be-tween the cost and effectiveness In particular we can im-mediately stop the potential interference to other recordingactivities (eg recording voice using smart phones) by turn-ing off the single cBeacon rather than stopping distributeduBeacons one by one or awaken cBeacon for localizationwhen needed

43 Timing for DownconversionWeassign two different ultrasonic frequencies to the uBeaconand the cBeacon and they are deployed in different locationsThe receiver can detect a downconverted beacon only whentheir ultrasonic beacons arrive simultaneously at the receiverA natural question is how UPS+ synchronizes the arrivalsof two types of beacons To do so the cBeacon persistentlytransmits a continuous wave (CW) whereas the uBeaconsadvertise short beacons every a few milliseconds Fig 4 usesa toy example to explain the design Given that signals fromthe cBeacon arrive continuously downconversion occursonly when a uBeaconrsquos signal arrives at the receiver In thismanner timing is dependent only on the signal arrival ofthe uBeacon and irrelevant of the cBeacon Furthermorewe demonstrate the timing through an actual experimentin Fig 3(a) where the uBeacon advertises pulses at 40 kHzwhereas the cBeacon transmits a single-tone CW at 50 kHzWe observe an apparent pattern of pulses at 10 Hz everysecond similar to the uBeaconrsquos advertising pattern at 40kHz although the cBeacon never stops its signals

44 Trilateration with ToA in UPS+Suppose that the ith uBeacon is located at position Ui andtransmits a beacon at time ts The receiver detects the arrivalof this beacon at time ti Thus the beacon takes timinusts secondsto propagate in air We can compute the pseudo-range asfollows

c times (ti minus ts ) = c times tb + |Ui minus P | (5)

where c is the speed of sound P(x y z) is the receiverrsquos loca-tion and tb is the clock difference between the receiver andthe uBeacons All the uBeacons are assumed to be well syn-chronized in time and advertise beacons at time ts c ti andUi are known to the receiver hence the preceding equation

5

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

contains four unknowns including three coordinate vari-ables in P(x y z) and (tb + ts ) The receiverrsquos OS kernel mayintroduce additional internal delays due to multi-threadingThese delays can be counted to tb To solve P(x y z) thereceiver must receive beacons from at least 4 different uBea-cons This approach is called trilateration At a high levelUPS+rsquos trilateration has the following components

bull Estimation of ToA The first component estimates theToA of each beacon from recorded audio data in a complexenvironment (see sect5)

bull Multiple Beacon Access The second component identi-fies the source device of each beacon when multiple uBea-cons are present (see sect6)

bull Enhancement ofBeaconsThe third component enhancesthe SNRs of downconverted beacons at both transmitterand receiver sides (see sect7)

The subsequent sections elaborate these components

5 ESTIMATION OF TOAThe key to trilateration is the estimation of ToA when thebeacon arrives at themicrophone A naive solution is to allowuBeacons to transmit single-tone pulses Then the receivercan estimate ToA by detecting the existence of a signal at thedownconverted frequency as shown in Fig 3(a) Howeversuch an approach is typically challenged by backgroundnoise echoes and the notorious frequency selectivity Thespread spectrum technique is widely recognized as a goodsolution to conquer these challenges In this section wepresent the unique spread spectrum technique adopted inUPS+ and then introduce the estimation approach

51 Spreading Beacons with ChirpsA chirp is a sinusoidal signal with a frequency that increasesor decreases over time The chirp spread spectrum (CSS)is a spread spectrum technique that uses wideband linearfrequency chirps to modulate information CSS has beenproven to be resistant against noise and multipath fading inacoustic channels [24 36] Suppose the sampling rate is fsand t0 is the time that the first sample is obtained Then aperiodic chirp is defined as

S[k] = cos(2π (f0 +

12(k mod K)∆f )tk

)(6)

where tk = t0 + (kfs ) and f0 is the start frequency attime t0 The transmitter periodically sweeps the spectrum of[f0 f0 + K∆f ] The swept spectrum is divided by K equalintervals in the unit of samples On the receiver end ToAcan be computed by correlating the received signal with apredefined chirp template If the correlation spikes in the

Chirp Chirp) Chirp) Chirp

45kHz

60kHz

40kHz

cBeacon

uBeacon

5kHz

15kHz

Time

Freqeqncy

Freqeqncy

Time

AudibleSpectrum

Pulse

Transmitting side

t1 t2 t3 t4 t5

Receiving side

Pulse

Global offset

Local offset Local offset

Air

Fig 5 Detected dynamic chirp beacons The cBeacon transmitsperiodic chirps during 45 sim 60 kHz whereas the uBeacon transmitsa 300 ms pulse beacon at 40 kHz every 500 ms The receiver detectssegments of the downconverted chirps

middle of the reception then it indicates a match The posi-tion of the spike corresponds to the beginning of the chirp(ie ToA)

SpreadBeacon in theAirThe straightforward approachis to allow uBeacons to directly yield chirps However a chirpsignal typically sweeps a wide band (eg 5 kHz) which con-siderably exceeds a uBeaconrsquos bandwidth (ie 2 kHz) InUPS+ we allow the band-wider cBeacon to transmit chirpsLet Su and Sc denote the beacons transmitted by a uBeaconand the cBeacon respectively

Su [k] = Au cos(2π futk )Sc [k] = Ac cos(2π (fc + 1

2 (k mod K)∆f )tk )(7)

where the uBeacon advertises at fu and the cBeacon periodi-cally sweeps the spectrum within [fc fc + K∆f ] When theabove two equations are substituted into Eqn 4 we obtainthe downconverted beacon as follows

Sdarr[tk ] = Adarr cos(2π (fc minus fu +

12(k mod K)∆f )tk

)(8)

A comparison between Eqn 8 and Eqn 6 clearly indicatesthat the downconverted beacon is in the form of a chirpbut the sweeping shifts to the downconverted spectrumie [(fc minus fu ) (fc minus fu ) + K∆f ] To intuitively understandthis phenomenon we present the experimental results inFig 3(b) where the uBeacon advertises a continuous single-tone pulse at 40 kHz and the cBeacon transmits periodicchirps from 45 kHz to 60 kHz Consequently the receiverdetects intact periodic chirps between 5 sim 15 kHz Ourtechnical trick is that the spreading occurs in the air insteadof at the transmitting side and thereby no additional cost andbandwidth are required at uBeaconsDynamic Chirp Spreading Now let us consider what

happens when uBeacons periodically advertise pulse beaconsTo illustrate our basic idea we use the toy example in Fig 5

6

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

Fig 6 Correlation in the time domainThe above picture showsthe spectrum via short-time Fourier Transform (STFT) the bottompicture shows the correlation result The clear correlation peakscould be found exactly at the beginning of each chirp signal

where the cBeacon transmits chirps continuously but a uBea-con advertises short single-tone pulse beacons with spacesDownconversion occurs during the window when a pulseappears at the microphone thus the receiver can no longerdetect an intact chirp but only segments of it In the figurethe green solid lines denote the captured chirp segmentsThese segments are identical in time and bandwidth both ofwhich are determined by the pulse interval However theirstarting and ending frequencies may differ ie the spreadchirps are dynamic and depend on the propagation distanceWe call this spreading technique Dynamic Chirp SpectrumSpread (DCSS) Fig 3(c) shows the downconversion resultsunder the same settings as shown in Fig 3(b) except thatthe signals of uBeacon are changed to periodic pulses Thefigure indicates that we can only observe the chirp segmentswhere the unsampled parts are filled with dashed lines

52 Pinpointing Dynamic Chirp BeaconsThe conventional correlation method which uses a statictemplate fails in identifying dynamic chirp segments in thepresent case because the template is not static In particulartwo unknowns namely the global offset (denoted by Γ) andthe local offset (denoted by τ ) exist due to the dynamic condi-tion The global offset indicates the starting frequency of thecBeaconrsquos chirp when it first arrives at the microphone Thelocal offset indicates the starting frequency of the uBeaconrsquospulse when it first arrives at the microphone These twotypes of offset are annotated in Fig 5 Thus ToA correlationmust be performed in two dimensions as follows

(Γτ ) =(KNminusK )argmax(Γτ )=(00)

1K

τ+Ksumk=τ

S[k] middotG(Γτ k)

G(Γτ k) = cos(2π (fc minus fu + (Γ + 05(τ + k) mod K)∆f )tk )(9)

111hellip111 0

Preamble (30ms) ID (40ms) Guard interval (30ms)

uBeacon1 (100ms)

0 1 1 1 1 0 1 0

bit (5ms)

1

parity bit (5ms)

uBeacon1 uBeacon2 helliphelliphelliphellipTime

Fig 7 Multiple beacon access Beacons are advertised in a prede-fined order to eliminate mutual interference Each beacon containsan ID field to distinguish devices

where S is the received signal and N is the total numberof samples in the received signal N should be the doublewidth of the uBeaconrsquos pulse to ensure that at least onechirp segment is captured G is the dynamic chirp templateThe output of the objective function suggests a tuple whichenables the correlation to spike at the appropriate offsetFig 6 illustrates the correlation results over the sample datashown in Fig 3(c)

Optimization Solving the aforementioned objective func-tion requires performing N lowastK correlations which will be aburdensome and energy-consuming task for a mobile deviceWe notice that once a smart device holds its position alldownconverted segments will share the global Γ which iscaused by the signal propagation delay from the cBeaconThus Γ can be easily found by correlating an intact chirptemplate with the audio The correlation spikes at Γ whichallows the chirp template to exactly cover all segments Thenwe start to slide τ to determine the local offset This opti-mization process can reduce the number of correlation toN +K Note that although the global offset is involved in theabove equation it is actually not used in the trilateration

6 MULTIPLE BEACON ACCESSSo far we have discussed how UPS+ can estimate the ToA ofa beacon out from a single uBeacon This section discusseshow UPS+ distinguishes beacons from multiple uBeaconsie multiple beacon access

61 Time SynchronizationTrilateration requires all reference devices (ie uBeacons) tobe appropriately synchronized in time otherwise each un-synchronized uBeacon will introduce an unknown variableto the clock difference thereby rendering Eqn 5 unsolvableIn this regard we equip each device with a low-power WiFichipset and adopt IEEE 1588 PTP for time synchronizationThe classic Network Time Protocol (NTP) fails to satisfy ourdemand because it has an average delay of 10ms which leadsto a ranging error of 10ms times 340ms = 34m in UPS+ Bycontrast PTP can achieve a high accuracy up to 100micros [53]or produce a ranging error of 34 mm In UPS+ the receiver

7

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

0 10 20 30 40 50 60 70 80 90 100Time(ms)

-005

0

005

01Am

plitude

Preamble ID Guard

(a) Received audio signals

30 40 50 60 70 30 40 50 60Time(ms)

0

1

2

3

4

5

Amplitude

Acquired Decode

0 1 1 1 1 10 0

30 35 40 45 50 55 60 65 70

(b) Decoded ID

Fig 8 Illustration of beacon decoding (a) The whole beaconsignals including the guard interval and (b) zoom-in view of theID field after the removal of the chirp carrier

does not need to be synchronized with uBeacons because itsclock offset is modeled in Eqn 5 already

62 FrequencyTime Multiple AccessPrevious works typically use FDMA to encode different bea-cons However this method does not work in our scenariobecause the downconverted beacon may sweep any segmentof the chirp ( Fig 5) which is highly correlated with the loca-tion of the receiver To address this issue we adopt two-levelencoding strategies in UPS+

cBeaconEncodingUPS+ assigns different sweeping slopesto different cBeacons such that receivers can quickly locateitself at the room level Since a cBeacon can cover about50 times 50m2 area a single cBeacon is enough for each roomMultiple cBeacons are isolated through walls We furtherapply on-off keying (OOK) to encode the IDs of uBeaconscovered by the same cBeacon and schedule them in exclusivetime slots

uBeacon Encoding Fig 7 shows the encoding of pulsebeacons uBeacons advertise their beacons in a predefinedorder to avoid mutual interference Each beacon containstwo fields preamble and ID The preamble is composed of 30ms pulses The receiver uses the preamble to estimate ToAand to align at the ID field The ID field contains 8 bits eachof which has an interval of 5ms These bits are encoded withFM0 Since FM0 requires all bits to be flipped in the beginningit could avoid the emergence of a continuous 8 bit pulses (iepreamble-like ID) The last bit is reserved for parity checkingThus UPS+ can completely support 128 uBeacons that arecovered by a single cBeacon An additional 30ms guard blankis reserved at the end to avoid the interference from echoesDifferent uBeacons are scheduled to advertise in various

Trigger signalMCU Transmitter

circuitBurst signal

Ultrasound

12Vp 24Vp

Fig 9 Transducer driver circuit

slots Users are allowed to configure the scheduling basedon their practices eg allowing non-adjacent uBeacons toadvertise simultaneously to decrease delays or increasingguard intervals for low-duty cycle

Beacon Decoding The receiver initially seeks the bea-con preamble in the recorded audio data via correlation (see Eqn 9) Once the preamble is obtained the receivers canidentify all the parameters of the chirp from the preambleThen the audio data are multiplied with the chirp template(see Eqn 8) to remove the chirp-based carrier and acquirebaseband signals over the ID field Subsequently lsquo0rsquo or lsquo1rsquois decoded every 5 ms by determining if a transition occursduring each bit interval (ie Bi-Phase Space Coding FM0)Finally the receiver uses the decoded ID to determine thecorresponding uBeaconrsquos location Fig 8 illustrates an ex-ample of the received beacon signals from which the ID issuccessfully decoded

ToA Estimation SupposeM preambles are found in theaudio data Let Bi denote the starting position of the ith

preamble in the samples where i = 1 2 M B1 is selectedas the baseline and the ith beaconrsquos ToA (ie ti ) is calculatedas follows

ti = ((Bi minus B1) mod (100ms lowast fs )) fs (10)

The mod operation eliminates the scheduling delay whichis an integral multiple of 100 ms ie uBeacons are sched-uled every 100 ms We also assume that beacon propagationdoes not traverse a scheduling period or beyond 100ms times340ms = 34m This assumption is reasonable because abeacon propagating over 15 m becomes nearly undetectableThe receiver substitutes ti into Eqn 5 for the trilateration

7 ENHANCEMENT OF BEACONThe SNR of the second-order harmonics is weaker than thefundamental which affects the accuracy and the effectiverange To mitigate this issue we introduce multiple enhance-ment approaches at both transmitting and receiving sides

71 Boosting TransmissionFirstly we boost the transmission at beacon devices

Transmitter Circuit The conventional output circuitfor the transducer is powered directly from the 5 V supplyie standard input voltage We firstly use a boost converter

8

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(a) Background noise

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15Frequency(KHz)

-120

-110

-100

-90

-80

-70

-60

Pow

er(d

B)

Primary microphoneSecondary microphone

(b) Downconverted chirps

Fig 10 Audio PSD at two microphones The operating spec-trum of beacons is between 5 sim 15 kHz (a) Two microphonesrecord ambient noise with identical power levels and (b) two mi-crophones record similar chirps but with different power levels

to convert the 5 V supply to 12 V DC voltage Then wedesign a tricky driver circuit to raise and double the voltageAs Fig 9 shows we use an inverter gate to provide a 180phase-shifted signal to one arm of the driver The other armis driven by an in-phase signal As a result in spite of 12 Vpower input the current from two arms are constructivelysuperimposed at the output This design doubles the voltageswing at the output and provides about 24 V peak voltagewhich is almost twice than the conventional UPS

Transducer Array Using an array of transducers is aclassical solution to improve the transmitting power of ul-trasound This solution has been adopted in many previoussystems For example Cricket integrates 2 transducers on asingle beacon device achieving a range of up to 10m Similarto ours LipRead [35] utilizes the second-order harmonic forlong-range voice attack Although an array consisting of 61transducers is used in LipRead it only uses 5 of them foreach frequency segment and totally supports 6 segmentsThus it actually uses 5 transducers for a specific frequencyto achieve a maximum range of 30 ft (or 91 m)Boost from cBeacon The effective range of UPS+ is

longer than LipRead because the strength of a downcon-verted beacon (ie Adarr) depends not only on the transmittingpower of the uBeacon but also on that of the cBeacon (seeEqn 4) The transmitting power of the cBeacon is 10 000timesstronger than uBeacon and thereby can enhance the strengthof a beacon to a relatively higher level With this in mindwe integrate one transducer in our uBeacon prototype forenergy and cost saving However it is easy to extend thecurrent prototype to accommodate an array of transducers

72 Turbocharging ReceptionSecondly we design an enhanced algorithm that operatesat smart devices It attempts to draw beacons out of back-ground noise when the receiver is far away from the beacondevices We notice that the majority of modern smart devices(eg smartphones) have two microphones The primary mi-crophone is typically mounted on the bottom to ensure a

(a) Before (b) After

Fig 11 Turbocharging results The two figures show the beaconsignals before and after turbochargingdirect acoustic path from the mouth The secondary micro-phone is mounted on top of the device to capture voice witha lower sound pressure level They are separated by approxi-mately 10 cm and therefore receive acoustic signals withdifferent pressures This condition offers an opportunity toturbocharge downconverted beacons

721 Understanding TwoMicrophones To understand thedual-microphone mode we conduct experiments to observethe responses of two microphones installed in an iPhone Xas follows

Observation 1 First we use the two microphones torecord ambient noise (without beacon signals) for referenceThe power spectral densities (PSDs) of the two audio dataare shown in Fig 10(a) The signals are nearly identical orhomogeneous in terms of PSDs regardless of the locations ofthe two microphones This phenomenon is also confirmedin our other experiments which are conducted under noise-only conditions using the same mock-up phones but in acrowded place or a busy lecture room Homogeneity is un-derstandable The noise in our effective spectrum (5 sim 15kHz) typically originates from thermal noise and householdappliances These noises are ubiquitously present in a roomwithout definitive directions

Observation 2 Second we show the beacon-present PSDspicked up by the two microphones in Fig 10(b) The effectivespectrum shows that the power level of the secondary mi-crophone is considerably stronger than that of the primarymicrophone (20 dB higher) because our beacon devices aremounted on the ceiling Their signals must travel additionalcentimeters to arrive at the primary microphone comparedwith the secondary microphone thereby resulting in pres-sure difference The directivity of ultrasound is stronger thanthat of audible sound due to its shorter wavelength Suchdirectivity further magnifies the difference

722 Turbocharging Algorithm The above two observa-tions suggest that the PSDs of the two microphones are quitesimilar regarding background noise but become differen-tiable when beacons are present These two key observationsinspire us to design an enhancement algorithm at receiversIntuitively the downconverted beacons are chirp segmentswhose spectrums are dynamic that is the segment movesamong different frequency bins Moreover the beacons areadvertised every hundred of milliseconds Both conditions

9

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

provide idle windows to smart device to estimate the back-ground noise Specifically when the PSD difference of twomicrophones is less than a threshold the receiver starts to es-timate the PSD of noise by assuming that no beacon signal ispresent currently otherwise the receiver estimates the PSDof signals and appliesWiener filter to turbocharge the beaconsignals We omit the introduction of the filter due to spacelimitations and advise readers to refer to [51 54] for detailsFig 11 shows the effectiveness of the algorithm in an extremecase where the receiver is located 15 m away from the uBea-con The figure shows that the turbocharged chirp signalsbecome distinguishable although they are nearly drownedin background noise before filtering

8 IMPLEMENTATIONThis section presents the implementation of two types ofbeacon devices Their prototypes are shown in Fig 2

bulluBeacons The uBeacons have functions similar to thoseof devices designed in Cricket [20] A simple solution is todirectly apply their designs to our system However theCricket design was released ten years ago and many of itscomponents are already outdated Thus we have to redesignour own hardware The core function is to drive an on-boardultrasonic transducer (MA40S4S [58]) to speak at 40 kHz ata specific time point We also reserve an ultrasonic receiverfor testing which is not required for UPS+ The low costWiFi chip ESP32 [59] from Espressif systems is integrated forthe PTP protocol and online configuration We synchronizeuBeacons every 32s for energy saving where the beaconbroadcast is advised between 025s sim 64s Each uBeacon issupplied by two 5 V-batteries

bull cBeacon We use a general vector signal generator toproduce periodic ultrasonic chirps which are further trans-mitted to the air through an ultrasonic speaker (Vifa [60])The output power is tuned to 10 W It sweeps the 10 kHzband from 45 kHz to 55 kHz over 100ms ie 100 kHz persecond As a result the downconverted spectrum is between5 sim 15 kHz regarding to the 40 kHz of uBeacons We inten-tionally reserve the top 7 kHz bandwidth (ie 15 sim 22 kHz)to avoid the signal distortion due to the non-ideal transitionof the low-pass filters at microphonesNote that the above spectrum is only given as an exam-

ple setting Some pets (eg dogs) might be still sensitive tothe 40 sim 65 kHz We could raise the spectrum to a higherrange (eg 70 sim 80 kHz vs 65 kHz) once the relative down-converted spectrum falls into 5 sim 15 kHz In addition ourchirp signals are repeated every 70ms Such repetition couldintroduce a 14 Hz sound However 14 Hz sound belongsto subsonic wave and is far below than the lower bound ofthe human hearing system (ie 50 Hz) Thus users are notaware of its presence

9 EVALUATIONA total of 15 uBeacons are deployed in a meeting room withan area of 9 times 3m2 in our department The speed of soundis set to 34438 ms in our experiments The speed of soundis relevant to the temperature thus a calibration before theexperiments is performed To test smart devices as manyas possible and in a cross-platform manner we directly usea built-in recorder of a mobile OS (eg iOS or Android) torecord audio clips for post-processing in MATLAB We useuBeacons equipped with a single transducer by default andan iPhone X as the default receiver unless noted

91 Evaluation in Zero-DimensionWe start by qualitatively comparing UPS+ with state-of-the-art UPSs and evaluate the accuracy of ToA estimation inzero-dimension

911 Comparison with State-of-the-Art We compareUPS+with past UPSs from nine different perspectives as listed inTable 2 (1) Only a few UPSs use ultrasonic speakers as an-chors due to their high prices PC is an attempt in this direc-tion It achieves a good accuracy at an extremely higher cost(2) Unlike normal UPSs Dolphin deploys custom-made ultra-sonic receivers to locate transmitters but has a shorter range(ie 3 m) (3) ALPS designs an embedded ultrasonic speakerwhich can achieve the operating range of 40 m Due to thelack of hardware we present the mean accuracy in the tableonly (4) both BeepBeep and ApneaApp use the speakers inthe smart devices for ranging (5) TUPS is the UPS that weimplemented across uBeacons but operates at the fundamen-tal frequency TUPS orientates to ultrasonic receivers beingsimilar to the ALPS Dolphin and Cricket Thus we chooseTUPS as a benchmark baseline in our evaluation Particularlywe have the following observations (a) Previous UPSs useeither ultrasonic transducers or speakers whereas UPS+ is aunique system that uses the two components jointly to builda hybrid UPS (b) UPS+ achieves comparable accuracy andeffective range as transducer-built UPSs and meanwhile theunit cost is maintained at an acceptable level (c) UPS+ isthe unique system that operates at the ultrasonic spectrumbut remains compatible with current smart devices

We also list the results of other four typical RF-based solu-tions for comparison Usually RF-based solutions have longeroperating range (eg around 100m) but their accuracies arelimited to meter or sub-meter level RFID-enabled Tagorambehaves a good accuracy but requires the prior knowledgeof track Importantly UHF RFIDs are unavailable by smartdevices due to lack of UHF readers Bluetooth-enabled iBea-con might be the most widely used commercial localizationtechnology which was initiated by Apple Corp Our realexperiments show that the average accuracy of an iBeaconis around one meter WiTrack locates targets at cm-level

10

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

Table 2 Comparison to Past Ultrasonic UPSs

Scheme Type Accuracy Range CostNode Dim Spectrum Modulation Inaudible Compt1

Cricket [20] Transducer 10 cm 10 m 10 $ 2D 40 kHz Pulse Yes NoDolphin [29] Transducer 234 cm 3 m 10 $ 3D 20 sim 100 kHz DSSS Yes No

PC [24] Speaker 43 cm 50 m 50 $ 3D 19 sim 23 kHz Chirp No YesBeepBeep [28] Speaker 08 cm 5 m ndash 1D 2 minus 6 kHz Chirp No YesApneaApp [41] Speaker 2 cm 1 m ndash 1D 18 minus 20 kHz Chirp No Yes

ALPS [42] Transducer 161 cm 40 m gt 10 $ 3D 20 sim 215 kHz Chirp No NoTUPS Transducer 351 cm 8 m 10 $ 3D 40 kHz Pulse No NoUPS+ Hybrid 495 cm 6 m 5 $2 3D 40 sim 65 kHz DCSS Yes Yes

Tagoram [55] RFID 8 mm 12 m 01 $ 3D 820MHz ASK No NoArrayTrack [56] WiFi 57 cm 100 m ndash 3D 24 GHz FDMA No No

iBeacon [57] Bluetooth 1 m 150 m 25 $ 1D 24 GHz DCSS No YesWiTrack [18] ndash 13 cm ndash ndash 3D 546 sim 725 GHz FMCW No No

1 The column of lsquoComptrsquo indicates if the solution is compatible with todayrsquos smart devices2 The cost in UPS+ does not contain the price of cBeacon each of which cost about 60$ and is shared by multiple uBeacons

using RF reflections at the cost of almost 2 GHz bandwidthCompared against RF solutions our work offers a trade-offsolution between the practicality and the accuracy to thecurrently available smart devices and to meet the growingdemand on indoor high precision localization

912 Accuracy in ToA Estimation As ToA estimation isthe foundation of the trilateration we evaluate the accuracyof ToA estimation In the experiment we place seven dif-ferent uBeacons 2 m away from the receiver while holdingthe receiver at a static location (ie in zero-dimension) TheuBeacons advertises beacons every 5 s We record a 10 minaudio across these uBeacons Given that the distances be-tween the receiver and the uBeacons remain unchanged atall times the downconverted beacons should arrive exactlyevery 5 s We compute the ToA of the remaining beaconsby considering the ToA of the first beacon as the referenceThese relative times should be an integral multiple of 5 sFig 13 shows the errors of the estimated ToA across theseven uBeaconsThe result suggests that approximately 003 ms error oc-

curs in the estimation even if the receiver remains at itslocation mainly due to the synchronization time delay andthe additional time consumption for processing audio dataat the receiver eg moving data from an audio capture sys-tem to a memory or disk Such estimation delay will incura potential 003ms times 34438ms = 103 cm error in the sub-sequent ranging or localization results Nevertheless sucherror is rather stable because the maximum standard devi-ation is around 002 ms This experiment demonstrates thefeasibility of using downconverted beacons for localization

92 Evaluation in One-DimensionThen we evaluate the ability of UPS+ in one-dimensionthat is both the uBeacon and the cBeacon are fixed at their

positions The receiver is moved away from the uBeacon byfollowing a straight line Their distance is increased from40 cm to 800 cm with a step interval of 10cm We are inter-ested in determining UPS+rsquos performance as a function ofdistance

921 Accuracy in Ranging We initially investigateUPS+rsquosranging accuracy We compute the pseudo-ranges whenthe receiver is located at different positions Adopting thepseudo-range at 40 cm as the reference the displacementsrelative to the reference in other positions are calculated Theexperimental trials are conducted 20 times in each positionThe effective range is defined as the maximum distance whenthe median ranging error is less than 20 cm The rangingaccuracy is shown in Fig 12 which only shows the resultswhen the distance is less than 300 cm due to the space limitWe have the following findings

bull Single-transducer made uBeacon Firstly we equipa single transducer on the uBeacon UPS+ achieves a meanaccuracy of 485 cm and a standard deviation of 28 cm whenthe distance is less than 25 m However the mean errorincreases to 2261 cm when the distance is beyond 25 mThe effective range is about 3 m This is due to the facts thetransmitting power of a uBeacon is limited to 09 mW andthe adoption of second-order signals suffers from a largerattenuation compared with that of the first-order signalsBoth factors decrease the SNR of downconverted beaconsthereby reducing the accuracy when the receiver is awayfrom the uBeacon

bullThree-transducermadeuBeacon Secondly we equipthree transducers (ie UPS+++) on the uBeacon to form asimple ultrasonic array UPS+++ has a mean accuracy of36 cm when the distance is less than 3 m Correspondinglythe effective range is extended to 6 m This result almostreaches the range of TUPS which operates at the first-order

11

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 3000

10

20

30

40

50

Error(cm)

Distance(cm)

UPS+ UPS+++ PC

50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200012345678910

Fig 12 Ranging accuracy as a function of distance UPS+ integrates a single transducer on a uBeacon UPS+++ integrates threetransducers on a uBeacon PC refers to [24]

U1 U2 U3 U4 U5 U6 U7

0

005

01

015

02

Err

or(m

s)

Fig 13 ToA estimation

10-1 100 101 102

Error(cm)

0010203040506070809

1C

DF

wo-turbochargingw-turbocharging

Fig 14 Turbocharging

1 2 3 4 5 6Distance(m)

0

002

004

006

008

01

Bit

Err

or R

ate

Fig 15 Bit error rate

signal but is equipped with a single transducer The cost isonly increased 3 $ than TUPS This suggests that multipletransducers can indeed enhance the transmitting power ofultrasound

bull Speaker made beacon device Thirdly another sim-ple approach for lengthening the ranging distance is to in-crease the transmitting power similar to PC which uses theultrasonic speaker to transmit chirps at 19 minus 23 kHz Thetransmitting power is tuned at 10W which is 10 000times thatof our uBeacon Consequently the ranging accuracy remainsat 43 cm even if the receiver is located at a distance of 6 mIn summary UPS+ can achieve centimeter-level ranging

accuracy and works comparably with ultrasonic rangingwithin an effective range of approximately 3 m when uBea-cons are equipped with single-transducers We believe that3 m range can fulfill major demands in practice particu-larly when uBeacons are attached to ceilings or walls Ofcourse we can increase the ranging distance by enhancingthe transmitting power or integrating multiple transducersif necessary

922 Accuracy after Turbocharging We evaluate the ef-fect of the PSD-aware turbocharging algorithm at a smartreceiver equipped with two microphones Android smart-phones are allowed to select audios through two micro-phones by turning on the stereo mode We place a Huaweismart phone at a distance of 5 m Fig 14 shows the compar-isons of ranging accuracy with and without turbochargingThe median error without turbocharging is approximately2113 cm due to the confusion from background noise Intu-itively a 50-sample shift will incur a 01ms time shift or 27 cmerror in ranging Dozens of samples have been obscured in

this case To improve SNR we apply the turbocharging al-gorithm to ToA estimation Consequently the median errorrapidly drops below 9 cm which outperforms the result twiceThis is because the turbocharging can improve the energyof the correlation peak by thrice more than that withoutturbocharging

923 Accuracy in Decoding Subsequently we evaluatethe UPS+rsquos decoding ability We skip the identification test ofthe cBeacon because the results are always accurate regard-less of how the cBeaconrsquos sweeping slope is changed Themain reason for this condition is that transmitting signals ofthe cBeacon is relatively strong (10 000times that of the uBea-cons) which results in easy decoding Here we only focuson the decoding of uBeaconsrsquo IDs Fig 15 shows bit error rate(BER) as a function of the distance In contrast to rangingUPS+ exhibits a strong decoding ability BER remains below8 even when the receiver is moved to a distance of 5 m

A considerable difference exists between ranging and de-coding because decoding has a loose requirement for pream-ble alignment In UPS+ each bit has an interval of 5ms Ourstudy implies that a bit can still be decoded even when a mis-alignment of 2 ms exists (ie involving 40 of the samples)because energy accumulation across 60 of the samples for abit is sufficient for bit decision By contrast a misalignmentof 2 ms causes a 34435 times 0002 = 6887 cm error in rangingBackground noise can easily obscure a few samples in thepreamble

93 Evaluation in Two-DimensionThen we evaluate UPS+ in two dimensions with respect toits localization accuracy

12

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

0 5 10 15 20 25 30 35Error(cm)

0

02

04

06

08

1

CD

F

UPS+Traditional UPS

Fig 16 Localization accuracy

2 3 4 5 6 7 8Bandwidth(kHz)

0

5

10

15

20

25

30

35

40

Erro

r(cm

)

Fig 17 Error vs bandwidth

0 5 10 15 20 25 30 35 40Error(cm)

0

02

04

06

08

1

CD

F

Xiaomi 6XiPhone XiPhone 8Huawei P10Samsung Note 8

Fig 18 Accuracy vs Device

0 5 10 15 20 25 30Error(cm)

0

02

04

06

08

1

CDF

Silence

Music

Speech

Multipath

Fig 19 Accuracy vs environment

Transducer Wi-Fi0

20

40

60

80

100

Aver

age

Powe

r(mW

)

Fig 20 Energy vs components

0 10 20 30 40 50 60 70Synchronizing Cycle(s)

0

02

04

06

08

1

Aver

age

Powe

r (m

W) Broadcast Cycle = 1s

Broadcast Cycle = 2sBroadcast Cycle = 3sBroadcast Cycle = 4s

Fig 21 Energy vs cycle

931 Accuracy in Localization We compare UPS+rsquos lo-calization accuracy with that of TUPS which uses ultrasonicsensor operating at ultrasonic bands as its receiver We per-form 150 experimental trials In each trial the receiver isplaced randomly in the evaluation environment Fig 16 plotsthe CDF of the 2D localization error We observe the follow-ings

bull UPS+ achieves a median accuracy of 459 cm and a 90thpercentile accuracy of 1457 cm in 2D localization By con-trast the median and 90th percentile accuracy of the TUPSare 351 cm and 1513 cm respectively These results areconsistent with Dolphinrsquos implementation [29 39] whichalso reports a median localization accuracy of 3 cm Thisresult demonstrates that UPS+ can achieve nearly the sameaccuracy as the traditional UPS even if it works at the second-order harmonics Interestingly the 2D localization accuracyof UPS+ behaves slightly better than that under 1D condi-tion because trilateration in 2D uses the difference in ToAMany uncertain common variables (eg sync delay at uBea-cons or audio processing at the receiver) will be canceled outfrom the difference

bull UPS+ and Dolphin achieve more than 10times improvementfor the mean accuracy over Cricket [20 61] which reports amean accuracy of 30 cm This is due to the rapid developmentof hardware which provides considerably higher resolutionin sampling frequency compared with that in a decade agoIn summary adopting the second-order harmonic as lo-

calization media can achieve the same accuracy as that withthe first-order UPS+ even exceeds some past UPSs

932 Impacts of Parameters Next we evaluate the local-ization accuracy as a function of different system parameters

Impact of Bandwidth We evaluate UPS+rsquos localizationaccuracy as a function of bandwidth The cBeacon sweepsthe spectrum with a constant slope Thus increasing the

length of preambles of the uBeaconsrsquo beacons is equivalentto increasing the bandwidth of the downconverted beaconsFig 17 shows the impact of bandwidth on localization accu-racy The plot demonstrates that the accuracy monotonicallyimproves with increased bandwidth In particular if UPS+uses a 2 kHz bandwidth then the median error reaches 1313cm and rapidly drops below 10 cm for bandwidths larger than4 kHz This result is attributed to increased bandwidth whichprovides finer granularity in separating the LOS path fromechoes In our default setting 4 kHz is adopted to balancetime delay and accuracy

Impact of Receiver Diversity The nonlinearity effect ofa microphone system is a key in UPS+ for ultrasonic down-conversion A natural question is whether this effect can beapplied to different model receivers To this end we test 2Daccuracy across four typical smart-phones The nonlinear-ity effect of microphone system is a key in UPS+ for theultrasonic downconversion A natural question is whetherthis effect could work across different model receivers Tothis end we test the 2D accuracy across five mainstreamsmart phones A slight difference exists among these devicesThe errors of the iPhone seriesrsquos are worse than those ofthe others because their microphone systems are better insuppressing in nonlinearity Even so we observe that thenonlinearity effect still exists as a ubiquitous phenomenon

Impact of EnvironmentWe repeat experiments underdifferent environments such as playing music noisy lectureroom and narrow space where the space is full of echoesat different frequencies The results are shown in the top ofFig 19 We could see that there are about 5 cm changes inthe median accuracy The worst case happens when playingmusic where the ambient sound might contain some high-frequency items However UPS+ has a strong ability to fightoff the echoes in narrow space due to the DCSS technique

13

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

Microphone

Left AirPod

Right AirPodiPad 1 iPad 2 Trace

Fig 22 UPS+ in real-world applications (a) shows how UPS+enables finding AirPods (b) shows how UPS+ enables pairing iPads(c) shows how UPS+ enables tracking hand arm through an AppleWatch

933 Energy Consumption Finally we present the energyconsumption of a uBeacon across with respect to compo-nents and cycles respectively Fig 20 shows the accumulatedenergy of the two main components which are persistentlyconsumed during one second It can be seen that the WiFicommunication for the PTP sync consumes 4times energy thanthe transducer The results show that wireless transceiver isindeed an energy hog in smart devices which is consistentwith many previous reports Next we estimate the actualenergy consumption during one second regarding to thebroadcast cycle (for localization) and sync cycle (for clocksynchronization) The results are shown in Fig 21 It is clearthat the low-duty cycle would save much more energy espe-cially when considering PTP sync Specifically a uBeaconcan save 58 energy when adjusting its sync cycle from fiveseconds to one minute With respect to our current settings(ie 1 s sync and 3 s broadcast) each uBeacon can workabout 5 months If the PTP sync cycle is increased to 64 sthe life can be prolonged to 8 months

94 Evaluation in Three-DimensionFinally we vision three practical applications which can ben-efit from the high accuracy of UPS+ They qualitatively test3D localization accuracy in several real-world applicationsas shown in Fig 22

Finding AirPods This application shows to find a pairof AirPods (Ver 1) Each AirPod integrates two microphonesbut we are allowed to access the bottom one only Its sam-pling frequency is limited to 16 kHz For the better perfor-mance we correspondingly adjust the chirp of cBeacon toensure the downconverted spectrum falls into 2 sim 7 kHzPairing iPads UPS+ could provide fundamental location

service to VR-system This application shows to pair twoiPads once their distance is less than 10 cm

Tracking Hand-Arm We apply UPS+ into tracking ahumanrsquos hand-arm through an Apple Watch (Ver 3) Thewatchrsquos highest frequency is up to 22 kHz However wefind that the built-in voice recorder uses M4A format whichfilters the frequencies above 16 kHz This is also an important

reason why we set the downconverted spectrum to 5 sim 15kHz which can adapt to any kinds of formats

10 LIMITATIONS amp CONCLUSIONWe present UPS+ a system that enables previous UPSs towork in todayrsquos smart devices We use the nonlinearity effectas an ultrasonic downconversion approach to pull down thefrequencies of ultrasonic beacons To do so we invent twotypes of beacon devices uBeacon and cBeacon We believethat the implications of such a solution can pave the way forexciting new directions for the exploration of the acousticcommunityHowever the current prototype still has a couple of limi-

tations We outline them for future workRelatively short rangeAlthoughwe have demonstrated

that three-transducer made uBeacon can achieve an effec-tive range of 6 m the range is still far shorter than those ofRF-based solutions Actually the short operating range is acommon drawback of sound based localization system be-cause acoustic wave attenuates faster than electromagneticwave However the advantage of UPS+ is in its extremelyhigh accuracy Our solution is more competitive for accuracy-sensitive applications such as ARVR systems localizing tinyIoT devices indoor navigation and so on Another reasonthat our solution is limited in the range is that we engineerthe system using off-the-shelf ultrasonic transducers for sav-ing cost We can explore ultrasonic speakers like ALPS [42]to design long-ranged uBeacons if necessary

Pre-deployment UPS+ requires pre-deploying beacondevices Actually all indoor localization solutions must useanchor devices like uBeacons to establish the localizationcoordinate system which is an unavoidable step

Post-maintenance Since all uBeacon devices are battery-driven in UPS+ users are required to update batteries everya fewmonths Twomethods can be utilized to mitigate this is-sue First low-duty cycle algorithm is able to significantly re-duce the energy consumption as shown in the evaluation Forexample all uBeacons transit to low-powered silent statesand are waken up by the smart devices when used secondthe recent advances in the wirelessly charging also provide adirection to design battery-free uBeacons in our future work

ACKNOWLEDGMENTSThe research is supported by NSFC General Program (NO61572282) UGCECS (NO 25222917) Shenzhen Basic Re-search Schema (NO JCYJ20170818104855702) and AlibabaInnovative Research Program We thank all the anonymousreviewers and the shepherd Dr Rajalakshmi Nandakumarfor their valuable comments and helpful suggestions

14

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

REFERENCES[1] Y Shu K G Shin T He and J Chen ldquoLast-mile navigation using

smartphonesrdquo in Proc of ACM MobiCom 2015

[2] J Haverinen andA Kemppainen ldquoGlobal indoor self-localization basedon the ambient magnetic fieldrdquo Robotics and Autonomous Systemsvol 57 no 10 pp 1028ndash1035 2009

[3] C Zhang and X Zhang ldquoPulsar Towards ubiquitous visible lightlocalizationrdquo in Proc of ACM MobiCom 2017

[4] S Zhu and X Zhang ldquoEnabling high-precision visible light localizationin todayrsquos buildingsrdquo in Proc of ACM MobiSys 2017

[5] CWang HWu and N-F Tzeng ldquoRfid-based 3-d positioning schemesrdquoin Proc of IEEE INFOCOM 2007

[6] J Wang and D Katabi ldquoDude wherersquos my card Rfid positioningthat works with multipath and non-line of sightrdquo in Proc of ACMSIGCOMM 2013

[7] J Wang F Adib R Knepper D Katabi and D Rus ldquoRf-compass robotobject manipulation using rfidsrdquo in Proc of ACM MobiCom 2013

[8] Y Ma X Hui and E C Kan ldquo3d real-time indoor localization viabroadband nonlinear backscatter in passive devices with centimeterprecisionrdquo in Proc of ACM MobiCom 2016

[9] Y Ma N Selby and F Adib ldquoMinding the billions Ultra-widebandlocalization for deployed rfid tagsrdquo in Proc of ACM MobiCom 2017

[10] J Xiong and K Jamieson ldquoArraytrack A fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2013

[11] Y Chen D Lymberopoulos J Liu and B Priyantha ldquoFm-based indoorlocalizationrdquo in Proc of ACM MobiSys 2012

[12] P Bahl and V N Padmanabhan ldquoRadar An in-building rf-based userlocation and tracking systemrdquo in Proc of IEEE INFOCOM 2000

[13] K Chintalapudi A Padmanabha Iyer and V N Padmanabhan ldquoIndoorlocalization without the painrdquo in Proc of ACM MobiCom 2010

[14] T A Johnson and P Seeling ldquoLocalization using bluetooth devicenamesrdquo in Proc of ACM MobiHoc 2012

[15] M Azizyan I Constandache and R Roy Choudhury ldquoSurroundsensemobile phone localization via ambience fingerprintingrdquo in Proc ofACM MobiCom 2009

[16] W-T Tan M Baker B Lee and R Samadani ldquoThe sound of silencerdquoin Proc of ACM Sensys 2013

[17] Y Gao J Niu R Zhou and G Xing ldquoZifind Exploiting cross-technology interference signatures for energy-efficient indoor local-izationrdquo in Proc of IEEE INFOCOM 2013

[18] F Adib Z Kabelac D Katabi and R C Miller ldquo3d tracking via bodyradio reflectionsrdquo in NSDI vol 14 2014 pp 317ndash329

[19] ldquoMIT Project Oxygenrdquo httpoxygencsailmitedu

[20] N B Priyantha A Chakraborty and H Balakrishnan ldquoThe cricketlocation-support systemrdquo in Proc of ACM MobiCom 2000

[21] R Want A Hopper V Falcao and J Gibbons ldquoThe active badgelocation systemrdquo ACM Transactions on Information Systems vol 10no 1 pp 91ndash102 1992

[22] N B Priyantha A K Miu H Balakrishnan and S Teller ldquoThe cricketcompass for context-aware mobile applicationsrdquo in Proc of ACM Mo-biCom 2001

[23] A Smith H Balakrishnan M Goraczko and N Priyantha ldquoTrackingmoving devices with the cricket location systemrdquo in Proc of ACMMobiSys 2004

[24] P Lazik and A Rowe ldquoIndoor pseudo-ranging of mobile devices usingultrasonic chirpsrdquo in Proc of ACM Sensys 2012

[25] P Lazik N Rajagopal B Sinopoli and A Rowe ldquoUltrasonic timesynchronization and ranging on smartphonesrdquo in Proc of IEEE RTAS2015

[26] G Borriello A Liu T Offer C Palistrant and R Sharp ldquoWalruswireless acoustic locationwith room-level resolution using ultrasoundrdquoin Proc of ACM MobiSys 2005

[27] S P Tarzia P A Dinda R P Dick and G Memik ldquoIndoor localizationwithout infrastructure using the acoustic background spectrumrdquo inProc of ACM MobiSys 2011

[28] C Peng G Shen Y Zhang Y Li and K Tan ldquoBeepbeep a high accu-racy acoustic ranging system using cots mobile devicesrdquo in Proc ofACM Sensys 2007

[29] M Hazas and A Ward ldquoA novel broadband ultrasonic location systemrdquoin Proc of ACM Ubicomp 2002

[30] M McCarthy P Duff H L Muller and C Randell ldquoAccessible ultra-sonic positioningrdquo IEEE Pervasive Computing vol 5 no 4 pp 86ndash932006

[31] J Yang S Sidhom G Chandrasekaran T Vu H Liu N Cecan Y ChenM Gruteser and R P Martin ldquoDetecting driver phone use leveragingcar speakersrdquo in Proc of ACM MobiCom 2011

[32] A Mandal C V Lopes T Givargis A Haghighat R Jurdak andP Baldi ldquoBeep 3d indoor positioning using audible soundrdquo in Proc ofIEEE CCNC 2005

[33] G Zhang C Yan X Ji T Zhang T Zhang and W Xu ldquoDolphinattackInaudible voice commandsrdquo in Proc of ACM SIGSAC 2017

[34] N Roy H Hassanieh and R Roy Choudhury ldquoBackdoor Makingmicrophones hear inaudible soundsrdquo in Proc of ACM MobiSys 2017

[35] N Roy S Shen H Hassanieh and R R Choudhury ldquoInaudible voicecommands The long-range attack and defenserdquo in Proc of USENIXNSDI 2018

[36] H Lee T H Kim J W Choi and S Choi ldquoChirp signal-based aerialacoustic communication for smart devicesrdquo in Proc of IEEE INFOCOM2015

[37] Q Lin L Yang and Y Liu ldquoTagscreen Synchronizing social televisionsthrough hidden sound markersrdquo in Proc of IEEE INFOCOM 2017

[38] A Harter A Hopper P Steggles A Ward and P Webster ldquoTheanatomy of a context-aware applicationrdquo Wireless Networks vol 8no 23 pp 187ndash197 2002

[39] M Hazas and A Ward ldquoA high performance privacy-oriented locationsystemrdquo in Proc of IEEE PerCom 2003

[40] V Filonenko C Cullen and J Carswell ldquoInvestigating ultrasonic posi-tioning on mobile phonesrdquo in Proc of IEEE IPIN 2010

15

[41] R Nandakumar S Gollakota and N Watson ldquoContactless sleep apneadetection on smartphonesrdquo in Proc of ACM MobiSys 2015

[42] P Lazik N Rajagopal O Shih B Sinopoli and A Rowe ldquoAlps Abluetooth and ultrasound platform for mapping and localizationrdquo inProc of ACM Sensys 2015

[43] M Rossi J Seiter O Amft S Buchmeier and G Troumlster ldquoRoomsensean indoor positioning system for smartphones using active soundprobingrdquo in Proc of ACM AH 2013

[44] K Lorincz and MWelsh ldquoMotetrack A robust decentralized approachto rf-based location trackingrdquo in International Symposium on Location-and Context-Awareness Springer 2005 pp 63ndash82

[45] H Lu W Pan N D Lane T Choudhury and A T Campbell ldquoSound-sense scalable sound sensing for people-centric applications on mobilephonesrdquo in Proc of ACM MobiSys 2009

[46] Y-C Tung and K G Shin ldquoEchotag Accurate infrastructure-freeindoor location tagging with smartphonesrdquo in Proc of ACM MobiCom2015

[47] A A Panchpor S Shue and J M Conrad ldquoA survey of methods for mo-bile robot localization and mapping in dynamic indoor environmentsrdquoin Proc of IEEE SPACES 2018

[48] M Jeub C Herglotz C Nelke C Beaugeant and P Vary ldquoNoisereduction for dual-microphone mobile phones exploiting power leveldifferencesrdquo in Proc of IEEE ICASSP 2012

[49] Y-Y Chen and J-H Zhang ldquoBackground noise reduction design fordual microphone cellular phones Robust approachrdquo IEEEACM Trans-actions on Audio Speech and Language Processing vol 25 no 4 pp852ndash862 2017

[50] R Martin ldquoNoise power spectral density estimation based on optimalsmoothing and minimum statisticsrdquo IEEE Transactions on speech and

audio processing vol 9 no 5 pp 504ndash512 2001

[51] M Jeub C Nelke H Kruumlger C Beaugeant and P Vary ldquoRobust dual-channel noise power spectral density estimationrdquo in Proc of IEEEEURASIP 2011

[52] P Horowitz and W Hill The art of electronics Cambridge Univ Press1989

[53] ldquoTI PTP Application Reportrdquo httpwwwticomlitansnla098asnla098apdf

[54] R G Brown P Y Hwang et al Introduction to random signals andapplied Kalman filtering Wiley New York 1992 vol 3

[55] L Yang Y Chen X-Y Li C Xiao M Li and Y Liu ldquoTagoram real-timetracking of mobile rfid tags to high precision using cots devicesrdquo inProc of ACM MobiCom 2014

[56] J Xiong and K Jamieson ldquoArraytrack a fine-grained indoor locationsystemrdquo in Proc of USENIX NSDI 2012

[57] ldquoBluetooth iBeaconrdquo httpsestimotecom

[58] ldquoTransducer Datasheetrdquo httpswwwmuratacomproductsproductdetailpartno=MA40S4S

[59] ldquoESP32-DevKitCrdquo httpswwwespressifcomenproductshardwareesp32-devkitcoverview

[60] ldquoUltrasonic Dynamic Speaker Vifardquo httpwwwavisoftcomusgvifahtm

[61] A K L Miu ldquoDesign and implementation of an indoor mobile naviga-tion systemrdquo PhD dissertation Massachusetts Institute of Technology2002

16