Sensing by the people, for the people & of the people

Anish Arora

April 21, 2008

Anish AroraAnish Arora

April 21, 2008April 21, 2008

The Samraksh Company

2

Overview

This is a people-centric perspective

Based on findings from recent experiments

−

OSU cellphone-mote based sensing

−

OSU-AFRL federated sensing

Plus we’ll look ahead at people-sensing needs

And argue the need for low-cost, low-power,

yet (currently lacking) information-rich sensing

3

Portable Cellphone-Mote

Motorola E680i/g & ROKR E2 phones

Intel PSI (Phone System Interface) Mote with accelerometer sensor

J2ME App(phone GUI)

TinyOS App(PSI mote)

SerialForwarder

(Linux)

Record management

system

a hack!

MMC/SDslotMessage

Transport(Linux)

Socket

+ =

4

TinyOS ported to PSI Mote

TinyOS-PSI interface for

MSP board

CC2420 board

GPIO (e.g. LEDs) ports

UART (including SPI) ports

ADC (e.g. accelerometer) ports

Reprogramming tools for Linux (host)

Serial forwarder for Linux

Available from TinyOS Sourceforgeor contact Lifeng Sang, sangl@cse.ohio-state.edu

5

building sensors

anchor nodes

mobile nodes

LSAP

PeopleNet: In-building Cellphone-Mote Network

Multi-hop network across cellphone-motes that are present

in CSE bldg (across all floors)

Rooted at a resource-rich building server, LSAP

anchor sensors

6

LSAP

Maintains

• Presence of Cellphone-Motes

• Routes (asymmetric)

• Interface with building sensors, other building LSAPs,

other networks

• Information exchange between cellphone-motes &

other sources

7

Cooperatively Tracking People Across Spaces

Exhibition people flow data collection: Most popular exhibit

Buddy location and messaging: Is Randy in his office?

already at the Rec Center? in which squash court is he?

8

PeopleNet is a Fabric: Users add Sensors & Apps

Room Occupancy Detectors: Conference Room, Squash Courtsusing PIR motion sensors

triggered by consistent activity over 2 minutes

heavily duty cycled sensing

On-line access useful; current reservation system imperfect

Dreese698

Dreese298

Free

Occupied

~24s ~13s

9

Elevator Sensor

Li/Kulathumani; idea not original: [ElevatorNet, Elson/Parker 2005]

10

Café Line Monitor

Our two building cafés have egregious lines at times

Idea being implemented:

• use radar sensor to estimate people count

• stream to LSAP

• query from PeopleNet

11

Dorothy’s Plants: Soil Moisture Monitoring

People often mis-(over-) water plants

Wireless sensors report on soil moisturestatus posted to LSAP/SensorMap

email or query

Written in DESAL [DESAL, WWSNA 2007]

experimental language to simplify sensor programming

This is an instance of more general Building Maintenance

service: take a photo or text a complaint from mobile

mote to Building Supervisor

12

In-building Camera Sensor

PIR-mote wirelessly trigger camera-motes

one outside building, one inside

Setup:

• 802.15.4 mote (Trio) with 1 active PIR

• 802.15.4 camera-mote

13

Federated Urban Sensing Scenarios

Satellites

UAV’s

Near-ground sensors

Low-power ground or in-building sensors

• Applications that use

multiple networks

• Information fusion

across layers

• Coordinated sensor

management

14

PIR

OSU-AFRL Federated Sensing Experiment

15

OSU-AFRL Experiment: Scenarios

PEOPLEVEHICLES

16

Persistent Coverage Scenario: Outdoor & In-building

Representative results: AFRL_SCENARIOS

19

Let’s Look Ahead: Broader Questions of People Interest

• Which was larger

the Girl Scout Sing Along

or the Anti-War March?

• National Park Service can’t say, even if it wanted to …

Data about public activities is under addressed

20

Site Activity

Detecting work stoppage

Detecting presence in

unsafe areas, e.g., below

cranes

21

Local Advertisements

22

Activity Heat Maps

Heat maps reduce information to enhance pattern recognition

Heat maps of human activity level could be useful

unusually low → ambush

unusually high → other problem

Useful for taxi services, malls, retailers, park vendors, outdoor festivals

23

What are They Doing?

Dancing?

Running?

Fighting?

Overactive?

24

Economic Studies

• Do youth still loiter?

• Are your parks being utilized?

is student center actually used?

are conferences rooms used when

reserved?

do employees use foosball table?

• If the courts are in use

does this imply safer streets?

25

Trade Show Traffic Flow

• Are there dead locations?adjust after show starts?

charge premium for hot spots?

• What is daily cycle as a function of location?

booths near lunch facilities have different daily cycle

• Physical adsTag visitors

Visitors get 30 second summary of the booth they’re looking at, delivered to an earpiece

26

Prediction

People centric applications will readily grow to exploit

substantial sensor deployment: scores in homes and

hundreds/more in offices

The value of this information will more often be in local

rather than global contexts

Problems of sensing cost, energy, and richness will matter

27

0 0.2 0.4 0.6 0.8 1 1.20

0.2

0.4

0.6

0.8

1

1.2

Single Sensor Coverage Area

Sen

sor C

ost

0 0.2 0.4 0.6 0.8 1 1.2 1.40

0.5

1

1.5

2

CPU Performance

CPU

Cos

t

Problem: Sensing Cost

Grosch’s First Law: CPU cost grows as square root of CPU

performance ⇒

buy the biggest computer you can afford

Sensor costs grow slower than coverage area

slide: courtesy of Ken Parker

28

Traditional Metric: Cost per Unit Coverage Area

Camera Towers:

• $100K, 8 km range

• ~200 sq. km per sensor

• $500 per sq. km

ExScal:

• $150 per sensor, 50m acoustic range

• ~133 nodes per sq. km

• $20K per sq. kmdoesn’t include Tier 2

29

New Metric: Cost per Unit Coverable Area

100

101

102

103

104

100

105

1010

1015

Terrain Range Limit in Meters

Dol

lars

per

km

2

$100K Sensor$10K Sensor$1K Sensor$100 Sensor

What if terrain, not sensor, limits range?

⇒ longer range sensor is just costlier

New performance measure:

• terrain range limit at which sensor is cost effective

• shorter range sensing can be better in urban settings

30

Problem: Communication Cost

Even when sensing range is limited, should communication

be high power, long range ?

• Communication range & timewindows sometimes be limited

• Wall power not always be available (acceptably or cheaply)

• Information not always needed at long distances

• Spectral efficiency of long range communications for many

sensors may not be high

⇒

Low-power, low-range sensor-communication has a role

31

Problem: Energy Cost

Mobility devices are readily rechargeable but…

cellphone users do complain if local sensing/comm energy

exceeds small % of energy budget (say 10%)

⇒

Challenge for mobile MAC is network discovery with almost

always off radio

discovery must be continuous and asynchronous

32

Node u

Node v

Listen SlotsBeacon Slot

……

……

Energy Efficient Asynchronous Discovery

3-state schedules: Beacon, Listen, Sleep

small packet size (<128 bytes) in low power radios

33

Optimal 3-state Wakeup Schedules

• Neighbor discovery may be Unidirectional or Mutual

• Optimal discovery exists that minimizes number of active slots in a given interval (frame)

• For example:

454443424140393837

363534333231302928

272625242322212019

181716151413121110

987654321

SlotBlock

BeaconWakeup

Frame

Optimal unidirectional discovery schedule for 45 slot frame

34

Optimal 3-state Wakeup Schedules

Similarly for mutual discovery

For example:

SlotBlock

Beacon

Wakeup

Frame

363534333231

302928272625

242322212019

181716151413

121110987

654321

Optimal Mutual Discovery Schedule for 36 slot frame

[Cao’07]

350 0.02 0.04 0.06 0.08 0.1 0.12 0.140

20

40

60

80

100

120

140

160

180

Duty cycle

Tim

e ta

ken

to d

isco

very

mut

ually

(s)

3-state schedule2-state schedule

Previous 802.11 work uses Active, Sleep schedules

[Tseng, Lai, 2005; Hou, 2002]

Beacon contention resolved in large packet size (>512 bytes)

Emulate 2-state schedule by random beaconing in active slots

Why Not 2-state Wakeup Schedules

Optimal 3-state vs. 2-state schedules:

requires energy

ensures deterministic discovery

same pattern for all frame sizes

21

36

Related Energy Problem: Dominant Receiver Power Consumption

Large portion of energy is consumed in receiver radio

* Batteries improving at only 12% a year

ExScal energy distribution:

receiver radio~2100 J/day

signal processing~60 J/day

everything else~8 J/day

Receiver Radio

Signal Processing

Other

37

Energy Efficiency Comparison

• Consider:Synchronous Blinking (S-MAC, T-MAC)

Long Preamble (B-MAC, WiseMAC, X-MAC)

Asynchronous Wake-up

Random Time-Spreading

• Traffic modelUniform random traffic

• Energy efficiency

=+

=∑ ∑∑ ∑

)( ji

ji

ji

RSM

EGoodput (Msgs Sent + Receive)

Total (Msgs Sent + Receive)

38

Optimal Energy Efficiency Comparison

10 15 20 25 30 35 40 45 500

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

The average number of nodes that interfere

The

ener

gy e

ffici

ency

Staggered On

Pseudo-random Staggered On

Long PreambleSynchronous Blinking

Asynchronous Wake-up

Random Time Spreading

39

O-MAC: Receiver Centric MAC for Mobility Devices

Transmitter centric MAC design:

transmitter implicitly knows receiver will wakeup during transmission

collision avoidance is transmitter driven (i.e., RTS-CTS, CCA)

Receiver centric MAC design:

receiver explicitly communicates its wakeup schedule to transmitter

collision avoidance is receiver driven (i.e., receivers use TDMA)

TransmitterReceiver

TransmitterReceiver

Transmitter

Receiver[ICNP06]

40

Staggered On Scheduling

Wakeup schedule:

• Only one receiver wakes up in the interference region at one time

• Scheduled globally to avoid receiver collision

• Each node knows its neighbor’s wakeup slot from global schedule

41

Pseudo-random Staggered On Scheduling

Frame

Slot Slot

Frame

SlotSlotSlot Slot Slot Slot

Listen

Transmit

Sleep

Frame

Slot Slot

Frame

SlotSlotSlot Slot Slot Slot

Sender

Receiver

Wakeup schedule:

• In each time frame, each node wakes up at a pseudo-random slot

• Each node knows neighbor’s wakeup slot by storing random seed

42

Optimal Energy Efficiency Comparison Revisited

10 15 20 25 30 35 40 45 500

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

The average number of nodes that interfere

The

ener

gy e

ffici

ency Staggered On

Pseudo-random Staggered OnLong Preamble

Synchronous BlinkingAsynchronous Wake-upRandom Time Spreading

• Receiver centric design achieves best energy efficiency

• Randomized staggered on still achieves comparable energy efficiency

43

O-MAC protocol design

• Based on: Pseudo-random Staggered On

• Core Protocol

Receiver centric synchronous communication

Asynchronous discovery despite:

−

Mobility

−

Link quality dynamics

−

Clock variations

Duty cycle adaptation

−

Dynamic traffic

Synchronous communication

AsynchronousDiscovery

Duty CycleAdaptation

Pseudo-RandomScheduler

44

Problem: Low Information Sensing

Extant point (temperature, pressure, humidity), pressure wave

(acoustic, seismic), & motion (PIR) sensors often inadequate

For sensing of people, richer spatio-temporal information neededcurrent sensors not sensitive or discriminating enough

Video imaging works, lower cost alternatives suffice for less

sophisticated user needs

Lot of work on networking, not enough on sensing

→

time to consider low-cost, low-power radar (esp. UWB),

e-field, electro-optical, chem, bio sensing

45

Pulsed Doppler Mote-Scale Radar

www.samraksh.com

• Range 10m (controllable to 1m)

• Coherent quadrature output (both I & Q)

• Responds to radial velocity 2.6cm/s -2.6m/s

• Range gate sharpness of 0.2m

• Interfaces with extant motes

• Low-cost

Apps:

• Velocity estimation

• Direction sensing

• Crowd estimation

• Robust motion detector…

46

Robust Motion Detection

48

Dealing with Clutter: Displacement Detection

• Bushes move, but don’t travel

• Targets often travel

• Instead of motion detection,

detect displacement

• Probably a bushor else not traveling

• Probably a targetat least not a bush

Tim

e

49

-500 0 500-500

-400

-300

-200

-100

0

100

200

300

400

In PhaseQ

uad

Phase and Frequency

Well known relations:

1. Frequency is rate of change

of phase, in radians per sec

2. Change in phase is target

displacement in wavelengths

Rotation about local center is

good measure of displacement

of dominant return

50

0 100 200 300 400 500 6001000

2000

3000

4000

5000

6000

0 100 200 300 400 500 600-20

0

20

40

60

80

Time in Sec

Wav

elen

ghts

0 100 200 300 400 500 600-10

0

10

20

30

40

Wav

elen

ghts

Results

• Person walking to & from radar 7 times

in alternating directions

• Around 530s is large motion of a bush

not as large as person nor as consistent

larger bush motions are possible

• This is phase rotation about

local center of rotation

bush’s motion nearly self-cancels

when person walks to or from radar,

net displacement is dramatic

can tell direction

51

In Conclusion

• Mobile-device based sensing opens up many relevant-to-

people applications

• Energy efficient networking of mobile devices is important

• The sensing itself needs to be low-power, low-cost, yet

information-rich

Pulsed Doppler Radar as examplar

• People should be able to readily add (not only sensors) but also their own apps