Hallway based automatic indoor floorplan construction using...

Hallway based Automatic Indoor Floorplan Constructionusing Room Fingerprints

Yifei Jiang†, Yun Xiang§, Xin Pan†, Kun Li� Qin Lv†, Robert P. Dick§, Li Shang�, Michael Hannigan∗

† Dept. of CS, �Dept. of ECEE, ∗Dept. of ME, University of Colorado Boulder, CO 80309 U.S.A.§ EECS Department, University of Michigan, Ann Arbor, MI 48109 U.S.A.† � ∗{yifei.jiang, xin.pan, kun.li, qin.lv, li.shang, hannigan}@colorado.edu,

§{xiangyun, dickrp}@eecs.umich.edu

ABSTRACTPeople spend approximately 70% of their time indoors.Understanding the indoor environments is thereforeimportant for a wide range of emerging mobile per-sonal and social applications. Knowledge of indoorfloorplans is often required by these applications. How-ever, indoor floorplans are either unavailable or obtain-ing them requires slow, tedious, and error-prone man-ual labor.

This paper describes an automatic indoor floorplanconstruction system. Leveraging Wi-Fi fingerprintsand user motion information, this system automaticallyconstructs floorplan via three key steps: (1) room adja-cency graph construction to determine which rooms areadjacent; (2) hallway layout learning to estimate roomsizes and order rooms along each hallway, and (3) forcedirected dilation to adjust room sizes and optimize theoverall floorplan accuracy. Deployment study in threebuildings with 189 rooms demonstrates high floorplanaccuracy. The system has been implemented as a mo-bile middleware, which allows emerging mobile appli-cations to generate, leverage, and share indoor floor-plans.

Author KeywordsIndoor floorplan; context sensing; indoor localization

ACM Classification KeywordsC.3.3 Special-Purpose and Application-Based Systems:Real-Time and Embedded Systems

INTRODUCTIONOn average, people spend approximately 70% of theirtime indoors [15], such as in offices, schools, and stores.New indoor mobile applications are being developedat a phenomenal rate, covering a wide range of indoor

Permission to make digital or hard copies of all or part of this work for per-sonal or classroom use is granted without fee provided that copies are notmade or distributed for profit or commercial advantage and that copies bearthis notice and the full citation on the first page. Copyrights for componentsof this work owned by others than ACM must be honored. Abstracting withcredit is permitted. To copy otherwise, or republish, to post on servers or toredistribute to lists, requires prior specific permission and/or a fee. Requestpermissions from [email protected]’13, September 8–12, 2013, Zurich, Switzerland.Copyright © 2013 ACM 978-1-4503-1770-2/13/09...$15.00.http://dx.doi.org/10.1145/2493432.2493470

personal and social scenarios. Many of these mobileapplications build upon indoor localization techniques,using location information to optimize and deliver on-the-fly personal and social services.

Indoor maps, i.e., indoor floorplans, are often requiredby indoor localization techniques to answer the follow-ing question “so... I’m here but where is here?” [11]. Sim-ply put, the floorplan is typically in the form of a build-ing blueprint, defining the structure and functionalityof a specific indoor environment. Given an indoor floor-plan, the personal and social functions supported by theindoor environment are then defined.

Acquiring indoor floorplan information is challenging– (1) many buildings do not have floorplans in easily-interpretable digital form; (2) even if available, buildingmanagers are often reluctant to share such informationwith the general public; furthermore, (3) buildings’ in-ternal structures and the corresponding functionalitiesoften evolve over time, making the original floorplansoutdated. Although it is possible to manually constructan indoor floorplan or correct an existing one, such pro-cess is slow, tedious, error prone, and difficult to scale.

This work tackles the problem of automated indoorfloorplan construction. To accurately construct an in-door floorplan, a set of indoor features are required, in-cluding (1) unique identification of individual rooms,(2) estimated room geometric information, e.g., lengthand width, and (3) geometric relationship amongrooms. Accurate determination of these indoor floor-plan features faces the following challenges:

• Heterogeneous indoor environments: An indoorenvironment consists of rooms with diverse sizes.Room connections through hallways also vary sig-nificantly. Such heterogeneous indoor environmentsmake accurate floorplan construction challenging.

• Noisy Wi-Fi fingerprints: Due to the complex mul-tipath propagation problem, Wi-Fi fingerprints ob-tained by mobile phones are dynamic and noisy. Theproblem is particularly challenging when using Wi-Fifingerprints to determine detailed floorplan features.

• Mobile crowd sourcing: Leveraging the built-in mo-tion sensors of the mobile phones carried by occu-

Session: Positioning I UbiComp’13, September 8–12, 2013, Zurich, Switzerland

315

pants, “crowd sourcing” offers a potentially highlyscalable solution for automatic floorplan construc-tion. The primary challenge is how to accuratelyextract stable and representative floorplan structurefrom diverse and random occupant motion patterns.

In this work, we propose an automatic indoor floorplanconstruction system. Leveraging Wi-Fi fingerprints anduser motion information collected via mobile crowdsourcing, the proposed system extracts indoor floorplanfeatures, including room identity, geometry, and inter-room geometrical relationship, and then construct theindoor floorplan automatically. To tackle the aforemen-tioned challenges, the proposed system uses the follow-ing main components.

• A room adjacency graph construction algorithm thatidentifies the adjacency of rooms and constructs aroom adjacency graph that is robust to the spatial biasof room fingerprints and Wi-Fi noise;

• A hallway layout learning algorithm that determinesthe room arrangement along each hallway, e.g., roomsizes and orders, using crowd-based motion sensingon smartphones; and

• A force directed dilation algorithm that adjusts the indi-vidual room structures globally to improve floorplanaccuracy.

The proposed system builds upon existing room lo-calization techniques [4, 6, 11]. The proposed systemhas been implemented and tested through deploymentin five different buildings with diverse floorplan struc-tures, including classrooms, research labs, libraries, of-fices, and shopping malls with 189 rooms in total. Thedeployment study shows that the proposed automaticindoor floorplan construction system yields an averageroom position accuracy of 91%, room area estimationerror of 33% and room geometric aspect ratio error of24%, requiring on average only 20 data points per loca-tion for the system to converge.

The rest of the paper is organized as follows. Westart with an overview of the system architecture, thenpresent the detailed technical design of the proposedsystem. Evaluation results are presented next. Aftersurveying related work, we conclude this paper.

SYSTEM OVERVIEWThis section formulates the problem of indoor floorplanconstruction, and provides an overview of the proposedsystem.

A floorplan graphically displays the dimensions andpositions of rooms inside a building. A floorplan in-cludes three types of information: the number of rooms,dimensions of rooms, and relative position of rooms. Inthis work, we focus on rectangular-shaped rooms typ-ically seen in most buildings. Each room can be rep-resented by the coordinates of its center point (e.g., xand y values on a 2-D plane), along with its width and

length information. The indoor floorplan constructionproblem can be defined as follows: Given a room set Rcontaining n distinct rooms on the same floor in a build-ing, determine the relative center coordinates, width,and length for each room ri(1 ≤ i ≤ n, ri ∈ R) in atwo-dimensional space.

In our system, the input data includes room Wi-Fi fin-gerprint ri for each room, passively collected motionsensor data (accelerometer and compass), and Wi-Fiscan data when users are walking in the building. Weassume that at least three Wi-Fi access points are visiblefrom most positions in each room and hallway. Thisassumption is valid for most public and commercialbuildings. A number of approaches exist for collectingroom Wi-Fi fingerprints, such as manual room finger-print collection [4], leveraging user feedback [8, 11] ,and automated learning techniques [6]. Our work lever-ages the outputs of the aforementioned approaches andwe assume that each room has been associated with aunique room ID and the corresponding Wi-Fi finger-print. Note that the room IDs are used for identifica-tion and are necessarily related to real room names ornumbers.

When users are walking, the motion sensor data andWi-Fi scan data are collected by our system running onusers’ smartphones. Note that all data are passivelycollected from users without explicit actions on users’part. Users do not even need to be aware of the collec-tion process. The accelerometer data are collected at a5 Hz frequency and used to determine if a user is walk-ing. When a user is walking, the compass and Wi-Fidata are collected at 5 Hz and 1 Hz frequencies, respec-tively. Note that this system only needs to be run in thetraining phase. Once the floorplan is constructed, thesystem no longer needs to be active, and the producedfloorplan will be provided to other indoor mobile appli-cations and services, thus saving smartphone energy.

Figure 1 illustrates the overall process of our indoorfloorplan construction solution. The system consists ofthree main components: (1) room adjacency graph con-struction, (2) hallway layout learning, and (3) force di-rected dilation.

Room Adjacency Graph Construction. This compo-nent identifies adjacencies among rooms and constructsa room adjacency graph. In a room adjacency graph, thevertices represent individual rooms and the edges be-tween vertices indicate adjacency between two rooms.We used k-means clustering to group Wi-Fi signals oftwo adjacent rooms into the same cluster. This approachis robust to Wi-Fi noise and unevenly distributed Wi-Fiscans within a room.

Hallway Layout Learning. Based on the room ad-jacency graph and users’ walking traces in hallways,the layout learning algorithm first identifies rooms andtheir orders along each hallway, and then assembles thehallways based on the similarity score of Wi-Fi signals.


316

65

1

4

23room ID with

Wi-Fi fingerprintsroom adjacency graph

construction6

5

1

42

3

65

14 2 3

65

14 2 3

room adjacency graph

hallway layoutlearningforce directed dilation

room layout along hallways

floorplan

accelerometer, compass, and Wi-Fi data while walking

Figure 1. Automatic indoor floorplan construction system overview.

This component includes a novel and accurate indoorturn detection algorithm, which leverages compass sen-sor data and a noise-robust room layout detection algo-rithm.

Force Directed Dilation. This step optimizes the over-all floorplan accuracy via global adjustment of the esti-mated room dimensions. We use a force directed dila-tion algorithm, in which the learned floorplan is definedas a mechanical system with springs between adjacentrooms, and the dimensions of rooms are automaticallyadjusted based on the forces among rooms.

ROOM ADJACENCY GRAPH CONSTRUCTIONRoom adjacency graph construction is a process to de-termine whether each pair of rooms is adjacent for agiven set of rooms, R. The adjacency information isrepresented using an undirected graph. For example,if two rooms are adjacent, the corresponding nodes ingraph are connected by an edge as shown in Figure 1.The adjacency graph is a fundamental information forfloorplan construction, and is used later in hallway lay-out learning and force directed dilation.

To determine the adjacency of two rooms, a straight-forward approach is to compare the similarity of theirfingerprints. However, such similarity can be affectedby the relative positions of two rooms, the type of wallbetween rooms, and the distribution of fingerprintswithin rooms. Yet, Wi-Fi signal noise increases the un-certainty. Moreover, this approach requires the defini-tion of a global threshold applicable to diverse roomconditions, which is very difficult to derive.

In this work, we propose a clustering based approach todetermine the relationship between pairs of rooms. It isbased on the observation that Wi-Fi signals in adjacentrooms are similar, especially for those signals collectednear the partition wall. Even though wall can decreaseWi-Fi signal strength, allowing differentiation of adja-cent rooms, noise can blur the boundary between tworooms. On the other hand, signals collected near thewall can be differentiated by extreme RSS (received sig-nal strength) values. The extreme RSS values for an AP(access point) are close to the minimum or maximumRSS value of that AP. Those signals are most likely at

the edges (corners) of rooms. By leveraging such Wi-Fi signals, we can cluster adjacent rooms into the samegroup.

Based on the above observations, for each room finger-print (containing a set of Wi-Fi scans collected in thatroom), we use the following procedure to remove Wi-Fi scans in which APs contain few extreme RSS values.First, for each access point api that can be observed ina room r, we extract its RSS value from the room fin-gerprint. Second, we normalize the RSS values usingminmax normalization to the range of [0, 1]. Given aWi-Fi scan w, where w = {rssap1 , rssap2 , ..., rssapm}, wecalculate the average normalized RSS value nrw as fol-lows:

nrw =m∑i=1

(rssapi)/m (api ∈ w). (1)

We remove all Wi-Fi scans whose nrw values satisfy0.2 ≤ nrw ≤ 0.8, and only keep Wi-Fi scans with moreextreme RSS values.

After removing those Wi-Fi scans, we run the k-meansclustering algorithm on the remaining Wi-Fi scans fromall rooms. Based on our evaluation results, we set kto 4 times the total number of rooms. For each clus-ter, we calculate the total number of Wi-Fi scans nscan,the number of unique rooms nroom, the number ofWi-Fi scans for each room nscan/room, and the rationscan/nroom as the threshold to determine if a roomshould be considered in this cluster. This is because ifa room only has a few Wi-Fi signals in a cluster, e.g.,nscan/room ≤ nscan/nroom, those signals might be noisecaused by environment changes or device heterogene-ity. Thus if there are two rooms in one cluster that eachsatisfies nscan/room ≥ nscan/nroom, these two rooms areclassified as adjacent rooms.

HALLWAY LAYOUT LEARNINGA hallway is defined as a straight or curved path in abuilding that is adjacent to rooms. The end of a hall-way can be a wall or an outlet. When a user walksthrough a hallway, he/she will pass all the rooms alongthe hallway sequentially. This indicates how rooms arearranged along the hallway.


317

Rooms vary by size and pairs of rooms can be posi-tioned in many possible orientations. Based on theseobservations, our layout learning technique includesthree sub-components: (1) hallway motion detection,which detects walking in straight hallways; (2) layoutlearning along hallway, which identifies the lengths ofrooms and sequences along the hallway; and (3) hallwayassembling, which is based on room overlap informationon intersecting hallways.

Hallway Motion DetectionThe goal of hallway motion detection is to identify useractivities, such as walking straight and making a turn.Our detection algorithm first distinguishes traversinghallways from walking within rooms by leveraging Wi-Fi signals while walking and room fingerprints. It thenidentifies activities in different hallways via turn detec-tion.

Hallway Traversal DetectionWe leverage the technique proposed by Jiang et al.[6]for walking detection due to its high detection accuracyand energy efficiency. In order to determine whethera user is walking in hallway, our system periodicallyscans Wi-Fi signals at around 1 Hz. When a user is mov-ing, the Wi-Fi RSS for access points change. However,as obstacles between phone and access points mainlyaffect Wi-Fi signal strengths, Wi-Fi signals within aroom are relative stable (even for moving phones) com-pared to those during hallway traversal.

Based on the observation, we determine whether amoving user is in a hallway or room by comparing twoconsecutive Wi-Fi scans wa and wb. Due to Wi-Fi sig-nal noise, the similarity score between two single Wi-Fiscans is quite unpredictable. In order to deal with thischallenge, we first convert the single Wi-Fi scan wi to aset of similarity scores dij for each room rj(rj ∈ R). dijis defined as [5]

dij =∏k

P (ngramk(wi)|rj)P (rj), (2)

where

ngramk(wi) = (apk, apk+1). (3)

That is, apk and apk+1 are two APs in Wi-Fi scan wi andare consecutive when ordered by RSS values. P (rj) isthe probability of room rj appearing in the system. Weset all P (rj) equal to 1.0 since rooms are treated equallyin our system.

The Wi-Fi scan wi is then converted to a vector Dwi=

di1, ..., din, where n is the total number of rooms in R.The similarity computation between wa and wb is con-verted to the similarity computation between Dwa

andDwb

. This conversion makes the similarity more reliablebecause dij is calculated using room fingerprint, whichcontains a set of Wi-Fi scans.

The similarity function between signal vector Dwa andDwb

is defined by Tanimoto Distance:

simi(Dwa, Dwb

)

=D(wa) ·D(wb)

|D(wa)|2 + |D(wv)|2 −D(wa) ·D(wb). (4)

We set a threshold τ . If simi(Dwa, Dwb

) ≤ τ , this sug-gests motion within a room. simi(Dwa

, Dwb) > τ sug-

gests motion within a hallway. Based on our experi-mental data, we set τ = 0.3.

Hallway Turn DetectionCompass-based turn detection [13] has been well stud-ied in outdoor environments. However, in indoor envi-ronments, the accuracy of digital compasses on smart-phones can be significantly affected by electrical cablesand devices, construction materials, and other metallicobjects nearby.

To deal with these challenges, we propose a techniquecombining accelerometer, compass, and Wi-Fi signals todetect turns in hallways. This is accomplished in threekey steps: (1) coarse-grained turning point detectionusing accelerometer-based and compass-based turn de-tection; (2) clustering of turning points using Wi-Fi sig-nals; and (3) fine-grained turning point detection whena user is close to a cluster of turning points. The intu-ition is that, we want to identify stable turning pointclusters and utilize fined-grained (i.e., more sensitive)turning point detection only when a user is close to aturning point cluster. Next, we describe the three stepsin detail.

Step 1: Coarse-grained turning point detection. Here, aturning point is detected whenever either accelerom-eter or compass detects a turn. Accelerometer-basedturn detection is inspired by the walking detection ap-proach [6], which assumes that when a user is walk-ing, his body is in oscillation. However, when the useris turning, the oscillation is broken. If m/mabs > 0.15and mabs > 300, we consider the user to be turning.Compass-based turn detection uses a 3 second longsliding window and calculates the mean value for eachwindow. If the change of two consecutive values islarger than τcompass (set to 30 in this work), the user isdetermined to be turning..

Step 2: Clustering of turning points. The individual turn-ing points detected using either accelerometer or com-pass can be noisy. Given the physical constrains of hall-ways, users’ turning points should be centered aroundcertain regions such as the end of a hallway or hallwayintersection. To identify these regions where turning islikely to occur, we leverage Wi-Fi scans collected at in-dividual turning points and use density-based cluster-ing to group Wi-Fi scans into clusters representing theseturning regions. The similarity of two Wi-Fi scans iscomputed based on RSS value using Equation 4. Thetwo parameters for density clustering algorithm, Epsand MinPts, are set to 0.8 and 15 respectively. Note


318

that although the similarity of two single Wi-Fi scans isnoisy, the density based clustering algorithm is robustto such noise and can produce stable Wi-Fi based turn-ing regions.

Step 3: Fine-grained turning point detection. Given the sta-ble turning regions, we can determine whether a useris close to any of these regions by calculating the simi-larity (Equation 2) between the current Wi-Fi scan andeach of the Wi-Fi clusters produced in Step 2. If the useris close to a turning region, we lower the thresholdsof accelerometer- and compass-based turning point de-tection. Specifically, m/mabs is set to 0.11 and τcompass

is set to 20. The lower thresholds make turning pointdetection more sensitive near the turning regions. Ifboth accelerometer- and compass-based approaches de-tect turning activity, the user is classified as turning.

Layout Learning Along HallwayGiven a specific hallways, there are two different typesof room layouts along this hallway: (a) rooms on oneside of the hallway only and (b) rooms on both sides ofhallway. Not only do we need to identify each hallwayand the rooms along the hallway, we also need to deter-mine the rooms’ orders and lengths along the hallway.

The hallway motion detection algorithm identifies theWi-Fi scan sequences collected in a particular hallway.Those sequences may cover a whole hallway or seg-ments of it. The hallway identification algorithm thenclusters those Wi-Fi sequences into groups so that eachgroup represents a unique hallway. All Wi-Fi sequencesin one group are collected in the same hallway. A traceis not required to cover the whole hallway from one endto the other. A user can start and end a trace at anypoints on a hallway. The system can learn the completehallway layout as long as the aggregated traces coverthe whole hallway.

For each hallway h, we get a group of Wi-Fi scan se-quences Sh and a room set s′h by comparing room fin-gerprints with Wi-Fi scan sequences. To identify roomorder and length of rooms along the hallway, we firstgroup rooms on the same side of hallway togetherbased on the room adjacency graph – if rooms are onthe same side of the hallway, their shortest path lengthto the group equals 1. Then for each side of the hallway,the following calculation is performed. For each Wi-Fiscan sequence sh ∈ Sh, sh = {w1, . . . , wj}, we calculatethe similarity between wi(1 ≤ i ≤ j) and each room ralong the hallway based on Equation 2. Then we canget a time series shown as real dots (blue) in Figure 2,where x is the time point at which wi was collected andy is the Wi-Fi similarity score. The time series tend tofollow Gaussian distribution, and fitting the time serieswith a Gaussian model can help reduce noise in the Wi-Fi similarity calculation..

We repeat this process for all rooms on the same sideof the hallway, producing results as shown in Figure 3.We compute the orders of rooms by compare the times-

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

0 5 10 15 20

sim

ilarity

sco

re

time (second)

modelreal

Figure 2. Hallway room similarity fitting.

tamps at the means of the Gaussian distributions. Thetime duration for crossing each room is also computedusing intersection of adjacent distributions. For exam-ple, in Figure 3, the start time ts2 of crossing room 2 isthe intersection of room1 (blue) and room2 (red) curvesand the end time te2 is the intersection of room2 (red)and room3(green) curves. The time duration of cross-ing room 2 is thus ts2 − te2. The room length alongthe hallway is computed by multiplying the durationwith average walking speed, which is set to be 1.4 m/s.The final room length is the average of room lengthscomputed from all sh in Sh. Using the average roomlength eliminates outliers caused by a few users whowalk slower or faster than others.

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

0 5 10 15 20 25 30

sim

ilarity

sco

re

time (second)

ts2

te2,ts3

te3 (half max)

room1room2room3

Figure 3. Hallway room order and size.

Hallway AssemblingWe have learned a set of hallways H and the infor-mation of room order and lengths along each hallwayh(h ∈ H). Hallway assembling aims to identify con-necting and/or intersecting hallways. The assemblingalgorithm starts with two hallways ha and hb (ha, hb ∈H). If ha and hb share common rooms, they are assem-bled as one of the structures in Figure 4 (a), (b), and (c).The choice of structure is determined by the positionsof the shared rooms in each hallway. For example, if theshared rooms are at one end of both hallways, the twohallways are assembled as the structure in Figure 4(a).We approximate the angle of two intersecting hallwaysas either 45◦ or 90◦. These two intersection types allowus to determine whether two hallways are perpendicu-lar, and in turn capture the geometric relations amongrooms.

In order to estimate the angle between two hallways,our system calculates the changes in compass readingswhen a user turns from one hallway to the other. Whenthe change is larger than a threshold τ , the angle of


319

(a) (b) (c)

hb

(d) (e)

hc

hc ha hc hd

ha

hb

Figure 4. Hallway assembling: Connecting multiple hallways.

the two hallways is classified as 90◦. Otherwise, it isclassified as 45◦. Due to noise resulting from magneticmaterial in buildings, a single estimation is unreliable.We use majority voting to determine the angle of twohallways by considering compass reading changes fromdifferent users and the same user at different times. Weevaluated this approach in four different buildings withcontrolled experiments. The results show that the hall-way angle classification accuracy is 93%.

Next, we pick another hallway hc that intersects withha or hb. The constraint of hallway intersection aloneis insufficient for the placement of hc. For example, inFigure 4(d), if hc intersects with hb at one end, hc can beplaced either above hb or below hb. In order to deter-mine that, we select a room rhc

from the side that doesnot intersect with hb and another two rooms rabove andrbelow which are above and below hb, e.g., two roomsat each side of ha. If the similarity between fingerprintrhc and rabove is greater than that of rhc and rbelow, hc isplaced above hb. Otherwise, it is placed below hb. Wedefine the similarity of two room fingerprints to be theaverage similarity of each Wi-Fi scan in one room to thefingerprint of the other room based on Equation 2.

Figure 4(e) shows one more scenario that needs to beaddressed in hallway assembling. Here, hd intersectswith ha and hc, but the length of hd is shorter than thedistance already learned between hc and hd. In thatcase, we dilate hallway hd to satisfy the conditions, andall rooms along hallway hd are dilated by the same pro-portion.

In some large public buildings, e.g., airports and shop-ping malls, a few hallways are curved. Figure 12 showsan example from a shopping mall. In order to deal withthese scenarios, our system calculates the hallway curveusing the hallway length and direction angle changesbetween the two ends of a hallway. The direction an-gle changes are calculated using the compass sensor onsmartphone. We use the average value over multipleusers in order to reduce the impact of magnetic noise.

FORCE DIRECTED DILATIONOur hallway layout learning algorithm is able to es-timate the room dimensions along each hallway (i.e.,room length) but not the width (or depth) of theserooms. To estimate the perpendicular dimensions, e.g.,room 105 and 150 shown in Figure 9(c), another tech-nique is required.

FrFsFs

(a) (b)

Fs

Figure 5. Spring system and the forces.

Algorithm 1 Force-Directed-Dilation(input G,D)1: input G; . graph after hallway learning &

assembling2: inputl D; . directions set: left, right, up, and down3: while G is not in equilibrium state do4: Compute Fvd on each v in d . v ∈ G and d ∈ D5: Fvjdk

= Max(Fvd)6: Stretch vj in direction dk with ∆ meters7: end while

One might consider using user motion within rooms toestimate room dimensions [1]. However, this approachcan be unreliable because it requires that users moveto all edges and corners of the room. In reality, usersare generally concentrated within part of a room. Thus,we propose to use a force directed dilation technique toestimate and automatically adjust the sizes of rooms.

Force directed algorithms [14, 3] have been applied tothe problem area of graph drawing. The idea is to placegraph nodes in 2D or 3D space with as few crossingedges as possible. The algorithm assigns an attractiveand repulsive force to each edge in a graph. The val-ues of the two forces are determined by the distancesbetween the nodes. Once the forces have been defined,the algorithms gradually adjust the positions of nodesuntil each node experiences net zero force.

Our floorplan optimization algorithm responds to at-tractive Fs and repulsive Fr forces, between rooms. Asshown in Figure 5, if two rooms are connected in thegraph, we assign a spring system between their centers.The force Fs exerted by a spring is defined as [3]. Notethat Fs also exists between hallways and rooms.

Fs(d) = c1 ∗ log(d/c2), (5)

where d is the length of the spring, and c1 and c2 areconstants. As shown in Figure 5(a), each room receivesthe same amount of force Fs but in opposite directions.We define Fr as the force generated by walls betweentwo adjacent rooms. Thus, if walls are not against eachother, as shown in Figure 5(a), the force Fr = 0. Other-wise, Fr is equal to the opposite force on the wall, e.g.,Fr = −Fs in Figure 5(b).

Having defined the attractive force Fs and repulsiveforce Fr, we use the force directed dilation algorithm,as sketched in Algorithm 1, to determine the dimen-sions of the rooms. The inputs of the algorithm arethe graph G and direction set D. G is learned in the


320

first two components and nodes have been placed aslearned/individual rooms. D represents the directionof force and consists of four directions: left, right, up,and down. A force F towards any other direction willbe decomposed into these four directions.

The equilibrium state of graph is defined as follows: (1)the net forces F on all vertices are zero or (2) no vertexin the graph can be dilated. If the graph is not in equilib-rium state, we compute the sum of forces for each ver-tex Fv and decompose it to each direction in D as Fvd.Then we find the largest force Fvd and dilate vertex v indirection d. Note that if there are multiple vertices withthe same largest force Fvd, all those vertexes will be di-lated. The algorithm repeats the above steps until thegraph is in an equilibrium state.

The reason to dilate a room with the largest force Fvd

is that when a vertex has a larger force, it means thatthe room has less similarity to the actual shape in thefloorplan. For example, in Figure 9(c), room 150 has thelargest force on the left direction, which is generated bythe attractive forces from room 105, 116 and 137. Thelarge force is caused by the deviation of room 105’s cur-rently assigned size from its actual size.

We assign higher priorities to rooms like 150 and 105 be-cause (1) adjusting those rooms will make the learnedfloorplan look more like the real one and (2) avoidingblock among rooms while dilation. For example, room116 in Figure 9(c) also receives force in the down direc-tion. If we change it first, it will block the dilation ofroom 150 in the following steps.

DEPLOYMENT STUDYThe system has been implemented as a mobile middle-ware on Android. A pilot indoor mobile applicationhas also been developed built upon the floorplan con-struction middleware. Nine users have been recruitedto conduct the deployment study at five different build-ing sites: classrooms, research facilities, libraries, shop-ping malls, and offices with a total of 189 unique rooms.The user study lasted two months.

During the user study, the participants installed ourmobile app on their personal Android phones and car-ried their phones normally without any change to theirdaily activities. The data collection function is automat-ically activated only when users are in the study build-ings in order to minimize the impact on smartphonebattery lifespan and protect participants’ privacy. Weimplemented a state-of-the-art room fingerprinting al-gorithm [6] to generate the Wi-Fi based room finger-prints that are required by the system as inputs. Thesystem automatically learned fingerprints for 91 uniquerooms using the algorithm. We manually collected fin-gerprints for the remaining 98 rooms in order to gen-erate complete floorplans. 11 phones of 6 types wereused. We collected more than 25 motion traces fromusers for each hallway.

0

0.2

0.4

0.6

0.8

1

1 2 3 4 5 6 7 8

pre

cis

ion

/ r

ecall

k / total number of rooms

recall precision

Figure 6. Quality of room adjacency graph with different k.

Room Adjacency Graph ConstructionWe first evaluate the proposed room adjacency graphconstruction algorithm, which provides fundamentalinformation for use in floorplan construction. We adoptthe precision-recall measure. More specifically, giventhe ground truth graph G and the produced graph G′

of our algorithm, the recall is the number of links in Gdivided by the sum of the number of links in both Gand G′s. The precision is the number of links in G′ di-vided by the sum of the number of links in both G andG′s. The precision-recall measure is evaluated underdifferent k-means clustering setting k. As shown in Fig-ure 6, as k increases, recall increases and precision de-creases. This is due to the fact that, with more clusters,the algorithm becomes more sensitive to the change ofadjacency between rooms. However, as the correspond-ing size of each cluster decreases, the classification pro-cess becomes less robust. The deployment study showsthat a k value equaling to 4 times of the total number ofrooms provides a good precision-recall tradeoff.

Motion-Based Traversal/Turn DetectionWe developed a motion detection technique for hall-way traversal detection and turn detection. Appropri-ate thresholds need to be experimentally determinedfor high-accuracy motion detection.

Figure 7 evaluates the quality of hallway traversal de-tection using different thresholds. Using precision-recall measure, the recall is the total duration of cor-rectly detected hallway traversals divided by the to-tal duration of hallway traversals and the precision isthe total duration of correctly detected hallway traver-sals divided by the total duration of detected hallwaytraversals. This study shows that the threshold settingsfor similarity between two consecutive Wi-Fi scans sig-nificantly affect the quality of motion detection. Smallthresholds result in missing hallway traversals. Largethresholds lead to misclassification of in-room walk-ing as hallway traversal. The deployment study showsthat a threshold of 0.3 provides a good precision-recalltradeoff.

Figure 8 evaluates the quality of hallway turn detec-tion using the precision-recall measure. The proposedmethod, which combines accelerometer, compass, andWi-Fi fingerprint data, is compared against the fol-


321

0

0.2

0.4

0.6

0.8

1

0.1 0.2 0.3 0.4 0.5 0.6

pre

cis

ion

/ r

ecall

threshold

recall precision

Figure 7. Traversal detection quality with different thresholds.

0

0.2

0.4

0.6

0.8

1

acc comp a+c ours

pre

cis

ion

/ r

eca

ll

recall precision

Figure 8. Turn detection quality comparison.

lowing three techniques, accelerometer-only, compass-only, and accelerometer-compass. It shows that,accelerometer-only based approach yields low recalland precision due to high dynamics and diversities ofhuman motion. Compass-only based approach yieldshigh precision but suffers from low recall when metal-lic objects are nearby. The combined accelerometer-compass approach further improves precision with thecost of low recall measure. On the other hand, leverag-ing the proposed Wi-Fi fingerprint-based turning pointdetection, our proposed solution offers both high preci-sion and recall.

Floorplan ConstructionFloorplans are the final output of the proposed system.We show four case studies of automatically constructedfloorplans, a classroom building, a research lab build-ing, a shopping mall, and a office builidng. The librarybuilding floorplan is not shown due to limited spaceand its similar structure to research lab building, but itsresults are included in the following quantitative anal-ysis results.

Figure 9 provides a step-by-step walk through of thefloorplan construction process for the classroom build-ing – (a) is the ground truth floorplan; (b) is the roomadjacency graph; (c) is the initial floorplan after hallwaylayout learning; and (d) is the final floorplan after forcedirected dilation. The results show that the automati-cally constructed floorplan offers excellent match to theground truth.

Figure 10 shows the research lab building case study,which also shows excellent match to the ground truth.In addition, detected floorplan mismatches are indi-cated. The following room pairs (indicated in red)had their positions swapped: 19–20, 40–41, and 49–

108 110 116 118

105 150

131 133 137 139

151

155

(a)

(c)

118

116110

108

150105131

133 137

139

151

155

(b)

108 110 116 118

105 150

131 133 137 139

151

155

108 110 116 118

105 150

131 133 137 139

151

155

(d)

Figure 9. Classroom building floorplan case study.

8 910 11

atrium

444546

atrium atrium

8 9

44 45 46

1920 20

19

40 4141 40 50 4949 50

Figure 10. Research lab building floorplan case study. Left side isthe original floorplan and right side is the learned one.

50. For rooms 44, 45, 46 (marked in blue), the widthis estimated incorrectly, because their widths are dif-ferent from the rest of the rooms along the same hall-way. Room 10 and 11 are elevators and emergency exits,which are not accessible to our study and were there-fore not detected and lead to overestimates of the sizesof rooms 8 and 9.

Figure 11 shows the office building case study, in whichsome rooms and hallways have irregular shapes. Theresult shows that our approach can still estimate roomsizes and relative positions with reasonable accuracyeven for irregular rooms and hallways. Figure 12 showsthe shopping mall building case study, which has amain curved hallway and rooms along the hallway. Thelearned floorplan indicates that the room relations arelearned correctly but the room size and shape are notexactly right. This is because we have no access to thehallway behind the stores and the room shape is mainlylearned through our force-directed dilation algorithminstead of the hallway layout learning algorithm.

To quantitatively measure the quality of the constructedfloorplan, the following three metrics are introduced:

• Position: A binary measurement for the relative posi-tion of two adjacent rooms with four directions: up,down, right, and left. Given two rooms, ra and rb, if


322

atrium

stair

Inaccessible Area

Figure 11. Office building floorplan case study. Left side is the orig-inal floorplan and right side is the learned one.

Figure 12. Shopping mall floorplan case study. Left side is the orig-inal floorplan and right side is the learned one.

any part of rb is on the left side of room ra’s left wall,we say that rb is on the left side of ra, similar conven-tions are used for other directions. Also the positionmay be a combination of two directions, e.g., up-left.The position accuracy is defined as the total numberof room pairs with correctly learned relative positiondivided by the total number of room pairs.

• Room area: The room length multiplied by width.Room area error is defined as the difference betweenlearned room area and ground truth room area di-vided by the ground truth room area.

• Aspect ratio: The room geometric length-width ratio,which reflect the shape of a room. Ratio error equalsto the difference of learned ratio and ground truth di-vided by real the ground truth.

Figure 13 shows the average position accuracy mea-sure as a function of number of motion traces per hall-way. Our system uses crowd-sourcing based learningmethod. In general, the more user data collected, thebetter its accuracy. Given the targeted accuracy mea-sure, a good learning method should converge quickly,and therefore have a low crowd-sourcing data require-ment. Given different numbers of motion traces col-lected per hallway, the corresponding floorplan accu-racies are shown in this figure. When 20 traces are col-lected in each hallway, the position accuracy convergesto 91%. Note that each motion trace can start and endat any point in a hallway.

Figure 14 shows the area error measure with and with-out the forced directed dilation. The results show thatforce directed dilation consistently improves room area

0

0.2

0.4

0.6

0.8

1

5 10 15 20 25

po

sitio

n a

ccu

racy

# of traces per hallway

Figure 13. Position measure.

0

0.2

0.4

0.6

0.8

1

5 10 15 20 25

roo

m a

rea e

rror


w/o FDSw/ FDS

Figure 14. Area measure.

estimation accuracy by 10%, achieving a stable errormeasure of 33% with 20 data points per location.

Figure 15 shows the average aspect ratio measure. Thisstudy shows that, force directed dilation consistentlyhelps reduce the error measure, converging to a 24% er-ror rate with 20 data points per location.

RELATED WORKIn this section, we survey related work spanning thefollowing areas: indoor mapping techniques, user be-havior sensing techniques, and Wi-Fi fingerprint basedindoor localization.

Floorplan constructionSimultaneous Localization and Mapping (SLAM) is atechnique to build a map in an unknown environment,while simultaneously tracking the current location [2].This technique is widely used in robotics. However, themap constructed in SLAM consists of a set of signifi-cant points, instead of floorplan. SmartSLAM appliesthe SLAM approach on the smartphones [12]. It uti-lizes the Wi-Fi signals collected when users are walkingto generate maps of a building. However, SmartSLAMfocuses on just hallway layout instead of the floorplanof the whole building. For example, it cannot derivethe room dimension and position information. Alzantotet al. have proposed a floorplan construction approach,CrowdInside, based on user motion traces in the build-ing [1].

Motion detectionThere is a rich literature on user behavior sensing tech-nologies [9, 7, 6, 10]. In this work, we implementedthe walking detection technology proposed by Jiang etal. [6] because it has good energy efficiency. Turn de-tection has been studied in previous work [7] using ac-celerometer and compass on smartphones. However,


323

0

0.2

0.4

0.6

0.8

1

5 10 15 20 25

aspe

ct

ratio e

rror


w/o FDSw/ FDS

Figure 15. Aspect ratio measure.

their technique is targeted for outdoor scenarios andthus does not consider the impact of indoor environ-ment on the accuracy of compass.

ComparisonsIn contrast with our work, CrowdInside [1] does notrequire Wi-Fi coverage and it works well for irregularrooms and hallways. However, CrowdInside relies onlandmarks (e.g., elevator and stairways), which are notalways available. Moreover, it requires that user tracescover all edges of a room in order to estimate its shape.This requirement may not be met in practice becauseusers typically concentrate in portions of rooms, e.g.,the desk area. In addition, the edges of rooms may beinaccessible to users, e.g., blocked by furniture. In ourfuture work, we would like to investigate leveragingtheir technique to further improve our system for irreg-ular rooms and hallways while maintaining the benefitof our approach that only rely on users’ motion in hall-ways.

We also compared our work with two state-of-the-artWi-Fi based room localization systems: ARIEL [6] andLiFS [16]. ARIEL [6] is a room fingerprinting sys-tem that automatically clusters Wi-Fi signals by room.In ARIEL, rooms have automatically assigned uniqueroom IDs and users can label rooms by their preference.It works well when the number of rooms is small foreach individual user, but a large number of room la-bels would be difficult for a user to memorize. In addi-tion, this approach only provides the room informationwithout any contextual information, e.g., room size andrelations to other rooms. Our system, built upon theARIEL system, can automatically construct a floorplanfor users to tackle the above problems.

LiFS [16] assumes a floorplan is available and it auto-matically maps Wi-Fi signals collected from users to theexisting floorplan. Their approach works well in build-ings whose floorplans are available. Otherwise substan-tial effort of manual floorplan construction may be re-quired. Our system deals with the case when floorplanis not available. LiFS is appropriate when floorplans areavailable.

CONCLUSIONSThis paper has presented an automatic floorplan con-struction system targeting indoor environments. Theproposed solution uses on-site Wi-Fi infrastructure. It

also uses information on the motion of occupants gath-ered via crowd sourcing. Our analysis and optimiza-tion techniques determine the identities and geometriesof individual rooms, as well as room adjacency infor-mation, and then construct an indoor floorplan throughhallway layout learning and force directed dilation. Thesystem functions for a variety of indoor environments,and is robust to Wi-Fi noise and variation in occupantmotion patterns. Deployment at three building siteswith 189 rooms yields an average room position ac-curacy of 91%, room area estimation error of 33% androom geometric aspect ratio error of 24%, using on av-erage only 20 data points per location for the system toconverge.

ACKNOWLEDGEMENTSThis work was supported in part by the National Sci-ence Foundation under awards CNS-0910995 and CNS-0910816.

REFERENCES1. M. Alzantot and M. Youssef. Crowdinside: automatic

construction of indoor floorplans. In SIGSPATIAL ’12, 2012.2. H. Durrant-Whyte and T. Bailey. Simultaneous localization and

mapping: part I. Robotics Automation Magazine, IEEE, 2006.3. P. Eades. A Heuristic for Graph Drawing. Congressus

Numerantium, 1984.4. A. Haeberlen, E. Flannery, A. M. Ladd, A. Rudys, D. S. Wallach,

and L. E. Kavraki. Practical robust localization over large-scale802.11 wireless networks. In MobiCom ’04, 2004.

5. Y. Jiang, K. Li, L. Tian, R. Piedrahita, X. Yun, O. Mansata, Q. Lv,R. P. Dick, M. Hannigan, and L. Shang. MAQS: A personalizedmobile sensing system for indoor air quality monitoring. InUbiComp ’11, 2011.

6. Y. Jiang, X. Pan, K. Li, Q. Lv, R. P. Dick, M. Hannigan, andL. Shang. ARIEL: Automatic Wi-Fi based room fingerprinting forindoor localization. In UbiComp ’12, 2012.

7. M. B. Kjærgaard, S. Bhattacharya, H. Blunck, and P. Nurmi.Energy-efficient trajectory tracking for mobile devices. InMobiSys ’11, 2011.

8. M. B. Kjærgaard, G. Treu, and C. Linnhoff-Popien. Zone–basedRSS Reporting for Location Fingerprinting. In Pervasive ’07, 2007.

9. M. B. Kjærgaard, M. Wirz, D. Roggen, and G. Troster. Detectingpedestrian flocks by fusion of multi-modal sensors in mobilephones. In UbiComp ’12, 2012.

10. F. Li, C. Zhao, G. Ding, J. Gong, C. Liu, and F. Zhao. A reliableand accurate indoor localization method using phone inertialsensors. In UbiComp ’12, 2012.

11. J.-G. Park, B. Charrow, D. Curtis, J. Battat, E. Minkov, J. Hicks,S. Teller, and J. Ledlie. Growing an organic indoor locationsystem. In MobiSys ’10, 2010.

12. H. Shin, Y. Chon, and H. Cha. Unsupervised construction of anindoor floor plan using a smartphone. Systems, Man, andCybernetics, Part C: Applications and Reviews, IEEE Transactions on,2012.

13. A. Thiagarajan, L. Ravindranath, H. Balakrishnan, S. Madden,and L. Girod. Accurate, low-energy trajectory mapping formobile devices. In NSDI’11, 2011.

14. W. T. Tutte. How to draw a graph. London Math, 1963.15. U.S. Environmental Protection Agency Green Building

Workgroup. Buildings and their impact on the environment: Astatistical summary, 2009.

16. Z. Yang, C. Wu, and Y. Liu. Locating in fingerprint space:wireless indoor localization with little human intervention. InMobiCom ’12, 2012.


324

Hallway based automatic indoor floorplan construction using...

Documents

Transcript of Hallway based automatic indoor floorplan construction using...