1

Infrastructure-free Collaborative IndoorPositioning Scheme for Time-critical Team

OperationsYoungtae Noh, Member, IEEE, Hirozumi Yamaguchi, Member, IEEE, and Uichin Lee, Member, IEEE

F

Abstract—Indoor localization is the key infrastructure for indoorlocation-aware applications. In this paper, we consider an emergencyscenario where a team of soldiers or first responders perform time-critical missions in a large and complex building. In particular, weconsider the case where infrastructure-based localization is not feasiblefor various reasons such as installation/management costs, a poweroutage, and terrorist attacks. We design a novel algorithm called thecollaborative indoor positioning scheme (CLIPS), which does not requireany pre-existing indoor infrastructure. Given that users are equippedwith a signal strength map for the intended area for reference, CLIPSuses this map to compare and extract a set of feasible positions from allpositions on the map when the device measures signal strength valuesat run time. Dead reckoning is then performed to remove invalid candi-date coordinates, eventually leading to only correct positions. The maindeparture from existing peer-assisted localization algorithms is that ourapproach does not require any infrastructure or manual configuration.We perform testbed experiments and extensive simulations, and ourresults verify that our proposed scheme converges to an accurate setof positions much faster than existing non-collaborative solutions.

Index Terms—Infrastructure-free Indoor Localization, Indoor Naviga-tion, Distributed Indoor Positioning Systems, Time-critical Team Oper-ations

1 INTRODUCTION

Location-based services basically aim to deliver customizedservices based on people’s locations. In the outdoor environ-ment, location-based services can be relatively easily deliveredwith well-known location services such as global positioningsystem (GPS) and cell-tower localization. However, such tech-nologies do not work well, and existing indoor location ser-vices usually require dedicated infrastructure indoors such asacoustic beacons [8], [9], [10], Wi-Fi access points (APs) [1],[2], [3], [4], [5], [6], [7], and radio-frequency identification(RFID) tags [11], [12], [6], [13].

In this paper, we consider a time-critical indoor scenar-ios with teams of soldiers or first responders performing

• Y. Noh is the corresponding author. He is with the Department of ComputerScience and Information Engineering, Inha University, Korea. E-mail:ytnoh@inha.ac.kr

• H. Yamaguchi is with the Department Information Networking, OsakaUniversity, Japan. E-mail: h-yamagu@ist.osaka-u.ac.jp

• U. Lee is the corresponding author. He is with the Department of Industrialand Systems Engineering, KAIST, Korea. E-mail: uclee@kaist.edu

emergency missions (e.g., firefighting, urban military, andsearch/rescue missions). If accurate and fast location servicesare available, the team members can easily access a region ofinterest for mission operations. However, in emergency sce-narios, infrastructure-based location services may face severalproblems. Deploying infrastructures such as acoustic, Wi-Fi,and Bluetooth beacons may not be feasible in time-criticalmissions. In practice, it might be neither economically feasiblenor practical to maintain such localization infrastructure foremergency missions in every building in advance. In someinstances, infrastructure might not even be available becauseof a power outage or terrorist attacks.

To summarize, our localization problem for time-criticalmission operations has several unique constraints: no existinginfrastructure can be used for localization (infrastructure-free)and team members cannot manually configure localizationschemes (configuration-free). The most widely-accepted solu-tion is inertial navigation (i.e., dead reckoning)—a user’s cur-rent position is tracked with inertial sensors (e.g., gyroscopesand accelerometers), by monitoring movement trajectories andpossibly matching a user’s current position in a map.

However, the well-known problems of inertial navigation in-clude large errors and slow convergence since indoor positionlandmarks are lacking; these problems may cause detrimentaleffect for time-critical indoor scenarios. When a team exploresa large and complex building, this problem becomes evenworse. For example, a team may enter an underground parkingstructure and then take an elevator to a locus of events on thefifth floor. The incremental rate in the position error could behigher than the rate at which the position can be fixed usingmap matching with turn detection [14], [15]. For accurate in-door localization, it has been well-accepted that we must makeuse of additional mechanisms such as manual configurationand infrastructure support. While peer-assisted localizationschemes collaboratively use peer measurements, they generallyrequire infrastructure support to improve localization accuracy.For example, Liu et al. [16] used peer-based acoustic rangingto improve the accuracy of radio frequency (RF) fingerprintbased positioning. In addition, peer nodes can collaborativelyperform simultaneous localization and mapping by combiningreceived signal strength (RSS) measurements from APs closeto the peers [17], [18], [19], [20], [21]. Unfortunately, existing

2

solutions are not particularly suitable for our emergency sce-nario because of its configuration-free and infrastructure-freerequirements.

In this paper, we propose a localization algorithm calledollaborative indoor positioning scheme , namely CLIPS. Ouralgorithm leverages peer-to-peer Wi-Fi beaconing and accu-rate smartphone dead reckoning. In CLIPS, obtained signalstrength (RSS) values are measured from other mobile peers.Given that a response team can obtain a floor-plan for safetyoperations, we can build a realistic RSS estimation mapin advance. For example, we can use a realistic wirelesssignal propagation model like ray-tracing. This estimation mapallows the team members to find a set of feasible coordinatesin which the estimated RSS values are matched with themeasured RSS values at run time.

To deal with wireless signal variations due to channelfading, we ensure that a position match can happen whenRSS value differences are within the threshold value. Thisprocess can initially produce many false candidates, but falsecandidates can be easily removed by leveraging dead reck-oning. That is, invalid candidates will lead quickly to deadends as a user moves along the corridors. This process willbe repeated such that each member’s real position can beestimated rapidly. In general, as the number of membersincreases, the convergence time decreases as well. In orderto conserve battery, Wi-Fi beaconing can be stopped whennodes? position fixes are entirely acquired; each user’s currentposition will be tracked by dead reckoning continues. In thenext processes, the search space (and convergence time) willbe drastically reduced, because it is enough to only considerthe locations closest to a node’s current location (e.g., withina fixed radius).

Concretely, this paper makes the following contributions:• We propose CLIPS, which uses collaborative Wi-Fi bea-

coning and smartphone dead reckoning. The CLIPS novelpositioning algorithm does not require any infrastructuresupport or manual configuration, which are the maindepartures from existing peer-assisted based localizationschemes [2], [16], [19], [20], [21]).

• We demonstrate the feasibility of CLIPS. The field testconfirms that CLIPS can accurately acquire position fixessignificantly faster than conventional approaches that arebased non-collaborative schemes; e.g., the required traveldistance to obtain an accurate fix took less than halfthat of a non-collaborative scheme (a more than 50%improvement in the considered scenarios).

• We further perform an extensive simulation study toinvestigate the impact of various parameters (e.g., theslack variable, number of nodes, mobility patterns, andlocation sharing) on the system performance under var-ious scenarios. Overall, we find similar trends to thosein the test-bed experiments. In particular, we confirmthat convergence speed exponentially decreases with theteam size, RSS matching contributes to a more than50% accuracy improvement, and sharing location updatessignificantly improves convergence speed.

The rest of this paper is structured as follows. We presentrelated work in Section 2. In Section 3, we overview CLIPS

and in Section 4, we review the core elements of CLIPS. Weillustrate our evaluation results based on the testbed experi-ments and simulations in Sections 5.1 and 5.2, respectively.We discuss the improvements and limitations of the currentwork in Section 6 and the paper is concluded in Section 7.

2 RELATED WORK

We first provide a general overview of indoor position-ing. CLIPS leverages multiple techniques, i.e., RF modeling,P2P communication (peer assisted), and dead reckoning, forinfrastructure-free instant localization. Thus, we detail eachof these key components and illustrate how CLIPS combinesthem.

2.1 Indoor positioning overviewIndoor positioning has been intensively investigated recentlyto provide accurate position information in indoor environ-ments, where GPS is not accessible. Existing indoor local-ization schemes can be classified into two categories, namelyinfrastructure-based and infrastructure-free systems. Previousindoor positioning schemes used specialized distance sens-ing devices, such as RFID tags [13], ultra-sound transmit-ters/receivers [22], [23], [8], and infrared IR beacons/receivers[10]. For example, Want et al. proposed Active Badge [10],where users wear a small badge that periodically transmits aunique infrared signal, and each room has networked sensorson the ceiling for accurate positioning. Similarly, ultrasound-based ranging is used in existing systems, e.g., Cricket [22],Active Bat [23], and WALRUS [8]. LANDMARC [13] usesRFIDs, where active tags periodically transmit signals, andreaders measure the signal strength for localization. Recently,researchers also showed that frequency-agile wireless net-works (i.e., in-band frequency hopping) can be used fordistance estimation. ToneTrack estimates the time-of-arrival(ToA) of a mobile node’s transmission at pairs of APs inthe network by analyzing the correlation between incomingsignals on different subcarriers [24]. It then compares the timedifferences of the ToA readings across pairs of APs to estimateand refine the mobile node’s location.

In contrast to these systems, which require additional in-frastructure installation, researchers have also proposed toleverage existing RF infrastructure, such as Wi-Fi and cellularcommunications. In RADAR [6], an RSS map is built a priori(using Wi-Fi signal strength vectors in the map). For a givensignal strength vector, the location of a node is then estimatedusing a nearest neighborhood matching algorithm in the map.In Horus [25], a stochastic description of the RSS map is usedfor maximum likelihood-based location estimation. Cellularsignals also have been used for localization [26], e.g., signal-strength fingerprints from multiple cell towers. However, if RFinfrastructure is uniformly deployed, it would be difficult tobuild a dense fingerprint map across an entire building. Tosolve this nonuniform map problem, Modellet [27] unifieda model-based and fingerprint-based approach such that aset of high-quality fingerprints are used to populate virtualfingerprints at positions without samples via a model basedapproach.

3

Recent commercial solutions tend to use multiple locationsources, such as GPS, cell towers, and Wi-Fi hotspots; e.g.,SkyHook [3], Google Latitude [5], and PlaceEngine [28].Recent studies have attempted to use other types of finger-prints. In SurroundSense [29], environment sensing data (e.g.,ambient sound and light conditions) were used for indoorplace detection. Likewise, given that an acoustic backgroundspectrum is fairly unique in each place, BatPhone [9] usesthis fingerprint for place recognition. Sextant [30] leveragesenvironmental landmarks such as store logos. It uses imageprocessing and databases to identify the landmarks and letssmartphones obtain relative position measurements. However,all these schemes require building fingerprint databases andmostly rely on infrastructure (e.g., Wi-Fi hotspots); thus, theyare not suitable for infrastructure-free instant localization. Amore complete survey of existing localization schemes can befound in recent surveys [31], [32].

2.2 Model Based Localization

Log-distance path loss (LDPL) as an RF propagation model,can be used to forecast RSS at various indoor locations [33],[2]. A clear benefit from this model is that it obviates theneed for an RSS map building process at the expense ofdecreased localization accuracy. Due to large diversities of RFpropagation characteristics, the model parameters would haveto be estimated particularly for each indoor space of interest.In TIX [34], knowing the transmit power and locations ofall Wi-Fi APs, the APs are modeled to measure the RSSof the beacons from nearby APs, and linear interpolationfinally estimates the RSS at every location in the indoor space.ARIADNE [36] also deploys sniffers at known locations, butemploys a more sophisticated ray-tracing model based ondetailed indoor maps and uses simulated annealing to estimateradio model parameters. Lim et al. [35] employed Wi-Fisniffers at known locations, and the LDPL model was usedto construct an RSS map. When there are peer mobile devicesnearby, EZ can collect RSS measurements of APs at the peerdevices to find LDPL radio model parameters—each node’sposition and radio parameters are treated as unknown variablesthat can be uniquely identified by using the RSS measurementsof APs at multiple nodes [2].

CUPID is based on the angle and distance estimationbetween an AP and a mobile node [37]. To accurately estimatedistance with the path loss model in an environment withmultipath fading, CUPID differentiates direct path (line-of-sight) signals from reflected path signals via the power delayprofile of the channel state information (CSI). Direct pathsignal strength information is then used to empirically findpath loss exponents. An angle-of-arrival (AoA) calculation thatis based on the phase difference between the signals that arriveat two antennas is also affected by multipath fading, becausedirect and reflected paths result in different angles. CUPIDmitigates this problem by leveraging a user’s mobility; i.e.,it computes distance traveled via dead reckoning, and thena cosine rule is used to estimate the change in the AoAs ofdirect path signals. Conventional localization requires multipleAPs for multilateration, but as shown earlier, RF infrastructure

may not be uniformly distributed. To mitigate this problem,Mariakakis et al. proposed a solution called SAIL that onlyuses a single AP [38]. When a user is mobile, it continuouslymeasures the distance between the client and AP. At thesame time, it estimates a user’s distance traveled via deadreckoning and the overall compass heading during that travel(this obviates the need for AoA estimation). Thus, a user’slocation can be geometrically determined using a single AP.

More recently, to cope with random signal fluctuations, Heet al. propose a novel indoor localization approach, named asWi-Dist, which fuses noisy wireless fingerprints with uncertainmutual distances given by their bounds [39]. Another methodalong similar lines, Wu et al. proposed the Automatic andContinuous radio Map Updating (AcMu) service, which takesadvantage of the static behavior of mobile devices [40]. Thissystem uses off-the-shelf mobile devices as movable referencepoints. These devices collect the up-to-date fingerprints whenthey are static at certain locations in real-time and are ac-curately pinpointed by a novel trajectory-matching algorithm.The underlying RSS relationship among neighboring locationsis stable over time with the help of newly collected data fromthe reference points.

Considering time-critical scenarios as discussed earlier, themain departure of CLIPS from previous solutions is that itleverages mobile Wi-Fi beacons (i.e., those of emergency teammembers) to determine a set of feasible coordinates fromall coordinates on an intended map, and then employs deadreckoning over the map to eliminate invalid candidate locationsand finally pinpoint the correct/real positions.

2.3 Peer-assisted LocalizationThe high density of smartphones/hand-held devices in publicspaces enables the peer-assisted localization of nearby peerdevices. Proximity sensing schemes, such as Hummingbird[41] and NearMe [42], provide relative localization of nodeswith an accuracy of 30 to 100 m. Hummingbird uses dedicatedRF ranging methods, whereas NearMe uses Wi-Fi fingerprintsto detect whether two nodes are within short range (if thereis at least one AP in common) or long range (if they canreached by hopping through APs with overlapping coverage)of each other. Bluetooth (e.g., BlueHoo [43]) and cellularsignals (e.g., People-Tones [44]) can be used for proximitysensing nearby devices, but its positioning accuracy is largelydependent on radio characteristics. Apple’s iBeacon uses Blue-tooth Low Energy, in which beacon nodes periodically sendunique IDs (say every 100 ms), for proximity sensing. Uponreceiving an ID beacon, nearby devices can look up theID to determine the device’s physical location or to triggerassociated services. Similar to NearMe, iBeacon categorizesproximity into three classes: 1) immediate (located within afew centimeters), 2) near (located within a couple of meters),and 3) far (located further than 10 m away). In contrast toWi-Fi and Bluetooth, acoustic ranging methods (e.g., BeepBeep [45]) can provide much higher accuracy than other radiobased methods. Recently, Banerjee et al. [46] proposed theVirtual Compass, where multiple radios are used to improvethe ranging accuracy (i.e., Wi-Fi and Bluetooth), and Vivaldi-based distributed localization is used to produce the relative

4

positions of nodes. Higuchi et al. [47] proposed “stop-and-go” collaborative localization that leverages accurate rangingmethods among mobile peers to estimate their moving states.As a result, more stationary nodes can be chosen as relativeanchors to improve accuracy. Liu et al. [16] used both Wi-Fibased positioning and acoustic ranging among peers such thatacoustic based relative positions are mapped into the absolutepositions by carefully referencing Wi-Fi based localizationresults. EZ can be considered a peer-assisted localizationbecause the RSS measurements of APs at the peer devices areused for localization [2]. Furthermore, as shown later, multiplenodes can collaboratively perform simultaneous localizationand mapping methods [17], [18], [19], [20], [21], which canbe also considered as peer-assisted localization. Our work canalso be considered as peer-assisted localization in that RSSmeasurement among peers are used for identifying feasiblecoordinates in the RF map, but the key difference fromearlier work is that our approach supports absolute localizationand uses collaborative dead reckoning, in which each nodeperforms map matching to eliminate infeasible coordinates andshares this information for fast convergence.

2.4 Robot Localization and Dead Reckoning

For the robot indoor navigation, a robot basically requiresan ability to determine its real location. Simple approachesprovision the robot with an indoor map that allows the robotto determine its real location by comparing its observedenvironment to the map (using, e.g., ultrasound or LADARsensors). Significant advances in robot mobility research in-clude Simultaneous Localization and Mapping (SLAM) [14],which allows a robot to build a map of the indoor environment(in terms of walls and other obstructions), and, at the sametime, determine its position based on the constructed map. Inorder to determine the distance having moved between mea-surement points, a robot typically uses an onboard odometer.For pedestrian navigation, Woodman et al. showed that the useof a foot-mounted inertial unit and particle filter can provideaccurate position estimation [48]. Lan and Shih [49] improvedthe accuracy of pedestrian dead reckoning by proposing anelaborate gait analysis technique for accurate step recognitionand step length estimation as well as a turn detection methodfor map matching based on gyroscope data. MapGENIE [50]is recent work on building indoor maps using pedestrian deadreckoning. It relies on crowdsourcing to collect the traces ofresidents. In addition, Sorour et al. [51] proposed joint indoorlocalization and radio map construction. The aim of this workis to reduce extensive deployment efforts for Wi-Fi radio mapconstruction. Our work differs from these earlier studies inthat we use P2P Wi-Fi fingerprinting and dead reckoning forfast location convergence in time-critical mission operations.

Recent studies have also explored the use of wirelessinfrastructure to better sense the environment [15]. Similarly,Martin et al. proposed a smartphone indoor localization thatuses only the smartphone sensors and integrates both onlineand offline RSS fingerprinting phases in a local smartphone[52]. Kim et al. [17] proposed a method for improving thelocalization accuracy of dead reckoning, where multiple nodes

collaboratively build the global RSS map, and drift errors arecorrected by referring to this map. SmartSLAM implementsSLAM to automatically construct an indoor floor-plan andradio fingerprint map using off-the-shelf smartphones [18].The cost of RSS map building can be significantly reducedby leveraging the fact that RSS fingerprints are geograph-ically connected as a user is moving, and thus, a high-dimensional fingerprint space can be formed. LiFS [21] usesmultidimensional scaling to map fingerprint space into two-dimensional space, which can be easily mapped onto a givenmap. UnLoc [19] leverages the fact that motion sensor andWi-Fi readings collected by multiple users can be clusteredto generate landmarks through which the drift errors of deadreckoning can be corrected. PiLoc [20] uses trajectory basedclustering to improve performance, i.e., trajectories are firstclustered based on AP similarity, path segments (turn and linevectors) are then clustered based on trajectories and RSS signalsimilarity, and path segments are finally merged to build floor-plans. These RSS map based methods cannot be used in ourscenarios because of their dependence on Wi-Fi infrastructure.Furthermore, landmark generation and trajectory clusteringbasically assume that there is sufficient sensor data collectedfrom mobile users, but this assumption does not hold in time-critical mission operations.

3 CLIPS OVERVIEWWe would like to emphasize that our work considers emer-gency indoor scenarios with first responders performing emer-gent missions (e.g., firefighting, urban military, or rescueoperations) or teams of soldiers. In this scenario, accurateand fast localization would facilitate team members to easilynavigate a region of interest, and sharing situational-awarenesswould finally lead to successful mission completion. However,we recognize there are limits to the accuracy of RSS-basedmap mapping—it does not deterministically identify exactlocations, although we have introduced a slack variable toaccommodate RSS variance and estimation errors and producean acceptable number of feasible coordinates/positions. Be-cause of this design choice, we need an additional componentthat can uniquely lead to a user’s real location/position fromthe feasible coordinates. The most common solution is inertialnavigation (i.e., dead reckoning), which tracks a user’s currentposition by continuously monitoring heading changes anddistance traveled with inertial sensors (e.g., gyroscopes andaccelerometers), and finally leads to current position of theuser in the map.

We assume that a team has obtained a floor-plan in advance(an example is shown in Figure 1(a)). It is preprocessed tobuild an overlay of an M ×N grid, as shown in Figure 1(b).The grid dimensions are determined by the granularity of thelocation information required for mission operations; we use a2×2 m grid in our field test. In addition, for all coordinates onthe floor, we calculate the path loss values from every othercoordinate on that floor, using a radio frequency (RF) ray-tracing model, and generate an RSS map of M × (N − 1)dimensions.

Figure 1(c) shows what would happen once a team offour members step on the floor in the intended building.

5

(a) Acquiring the floor-plan (b) Generating an RSS map using a ray-tracingbased RF simulation

55dB 70dB

90dB

(c) Measuring the RSS values to all reachablemembers via Wi-Fi beaconing (e.g., top-rightnode)

55dB 70dB

90dB

90dB 70dB 55dB

(d) Identifing the feasible coordinates of eachnode (e.g., one valid and five invalid coordinatesindicated by the white and black phones, respec-tively)

!

!!

!!

(e) Eliminating invalid candidates via dead reck-oning (e.g., all five invalid coordinates)

User 1 User 2

User 3

User 4

(f) Sharing the discovered position with otherteam members (via a cellular network or Wi-Fi) such that others can further remove invalidcoordinates

Fig. 1. CLIPS overview [53]

Each mobile node performs periodic Wi-Fi beaconing throughwhich the RSS values of the reachable team members can beobserved. As shown in Figure 1(d), these RSS values are thenmatched against the RSS map downloaded in advance, andthis generates a list of all feasible coordinates on the floor. Wepresent the detailed matching algorithm in Section 4.2. Afterobtaining these feasible coordinates, team members employsmartphone dead reckoning to eliminate all false positivesfrom the list. According to heading changes and distancetraveled on the map, each member applies the displacementon the list of all feasible coordinates. In the meantime, ifthe coordinates lead to a dead end, those invalid coordinatesare eliminated from the list. An example is shown in Figure1(e). This process is repeated until each member gets anunique coordinate from the list. As shown in Figure 1(f), onceeach member localizes herself correctly, she can broadcast herposition on the map to the remaining nodes such that she canhelp others to remove invalid coordinates on their lists quickly.This also allows each team member to know the position ofall other members, and helps to facilitate the team mission.

4 CLIPS SYSTEM DETAILS

In this Section, the main components of CLIPS will be pro-vided: (1) floor-plan pre-processing and RSS map generation,(2) feasible coordinate estimation, and (3) map matching viaindoor path tracking.

4.1 RSS Map Generation

CLIPS requires a floor-map for each site, which can beobtained from geographic information system (GIS) data. Weassume that a team under time-critical indoor operation canaccess a digital map from a service provider that recognizes

a given floor-plan image and builds a digital floor map—architectural floor-plan image recognition has been studied inthe field of GIS [54]. Similar services have recently beenlaunched, such as “Google Maps Floor Plans,” for indoorWireless LAN positioning, which obtain floor-plan imagesuploaded by users [55], and floor level information can beaccurately identified by using existing algorithms (e.g., FT-track [56]).

Next, the LDPL model as an RF propagation model isemployed to forecast RSS values for all coordinates on the in-tended floor-plan, and to obtain path loss data after simulation.Using this model reduces the number of RSS measurementssignificantly compared to RF fingerprinting schemes, albeit atthe expense of degraded localization accuracy. Given that RFpropagation characteristics vary widely (especially indoors),the model parameters would have to be estimated specificallyfor each indoor space of interest. For example, Anderson et al.presented measured data for a 2.5 GHz in-building partitionloss [57].

Ray-tracing has widely been used because ray-tracing mod-els consider the details of a place, such as thickness of walls,doors, and windows as well as desks and chairs to achieve highfidelity. In general, ray-tracing approximates radio propagationusing a finite number of isotropic rays emitted from a trans-mitting antenna by a ray imaging technique and can deliverhigh fidelity. In this technique, the transmitter is assumedto be reflected at each surface around it to produce imagetransmitters; the rays reflected to the receiver from the realtransmitter are considered to be direct paths from the mirrorimages of the true transmitter. Based on geometrical optics,each ray from the transmitter to the receiver can be exactlydetermined. However, the major drawback of such techniquesis their expensive computational complexity. Readers can referto the detailed ray-tracing techniques in [58], [59], [60], [61],

6

[62], [36]All models can be used in our extensive simulation. How-

ever, in order to pursue a reasonable balance amongst theeffort of floor map modeling, and accuracy, and simulationduration, we use the following modeling and simulation: weconsider all the solid lines in a floor-plan as walls, all thespaces surrounded by the solid lines as accessible spaces, andall windows are regarded as walls. With this indoor modeling,we use a simplified (reduced) 3-ray tracing in the exper-iments (provided by Wireless InSite software [63]), wherediffraction along obstacles is considered. In practice, however,the measured signal strengths tend to fluctuate because ofsmall-scale fading. This phenomenon can be mitigated byconducting multiple measurements to determine the averageRSS values, but a node might still suffer from short-termfluctuations (thereby removing the true position). Moreover,floor-plans that are partially outdated because of recent re-modeling, missing obstacles (e.g., furniture or refrigerators),human movements, and different Wi-Fi chipsets can also affectthe RSS values. To address these problems, we introduce aslack parameter α (usually ±13 dB in our settings) to reducethe sensitivity to RSS measurements on the overall systemperformance. We also believe that the calibration process ofWi-Fi chipsets will further opt-out false-positive positionsby decreasing the α value; however, this will be a part ofour future work. We justify this approach with test bed andsimulation results in Section 5.1.

4.2 Feasible Coordinate Estimation

Node j estimates the path loss value in this communication(denoted as m(i, j)) from each beacon log (i, j, ss)t of nodej, where ss is the RSS from node i at time t. To this end,we assume that transmission power and other factors, such assender antenna gain/loss and environmental noise, are commonand almost constant on all mobile nodes. We note that thesevalues might be hardware-dependent,1 but pre-measurementbefore localization (before a mission starts) is an effective wayto determine such values.

4.2.1 Positioning problem formulationIn this section, we focus on the positioning activity of node j.Given a floor-plan of fp with N grid points, let L : N2 → R+

denotes a path loss matrix among the N points, where element(u, v) ∈ N2 is the simulated path loss from points u to v. Inaddition, let Mj denotes the set of node j and the nodes fromwhich it received beacons at time t. Let m :Mj ×{j} → R+

denotes the set of path loss measurements of node j at timet, where m(i, j) is the measured path loss value from nodesi to j.

The node j’s positioning problem is to select the grid pointthat has the minimum distance from the true position of nodej . Our approach is to locate such a position by finding the

1. We introduce a slack parameter α and perform experiments to findemphatically correct value (shown in Section 5.1.2, ±13dB in our settings) toreduce the dependency as we cannot completely consider potential changesof floor-plans, missing obstacles (e.g., furniture, refrigerator, etc.), humanmovements, and different Wi-Fi chipsets to obtain RSS values.

positioning function p : Mj → N with the least “path lossmatching error” between m and L. The objective function isminimizing such path loss matching error as follows:

min∑i∈Mj

|m(i, j)− L(p(i), p(j))| (1)

We note that we may use different objective functions depend-ing on signal attenuation at the target site. For example, if thepath loss increases as the power of the distance, we may usethe following function instead of (1):

min∑i∈Mj

(m(i, j)− L(p(i), p(j)))γ (2)

where γ is the pass loss exponent (e.g., γ = 2 in the freespace attenuation model). More general propagation modelsexist. For indoor space, there is a well-known model in ITU-R P.1238-8 [64]. Assuming 2.4 GHz propagation on the samefloor with the reference distance 1 m, the P.1238-8 modelhas the form “30 log10 d − 39.60”. By transforming this log-based model into the exponent-based model, the coefficient“30” corresponds to the pass loss exponent γ = 3 in thepropagation model. We note that shadowing loss and mul-tipath fading are not usually considered in the general (i.e.,site-independent) propagation models because of their time-varying and situation-dependent features and we thereforeignore them in the objective function. Nevertheless, the pathloss characteristics are highly situation-dependent, and it is notstraightforward to choose the best-fit exponent. Consequently,we adopt the simplest form for (1) considering the simplicityof its calculation.

As addressed above, we need to accommodate measurementerrors caused by the fluctuation of RSS values. We shouldallow some deviation parameter α such that a measured pathloss l′ and a simulated path loss l are regarded as identical ifl′ ∈ [l − α, l + α]. The choice of an appropriate α is furtherinvestigated in Section 5.1. Using function z, where z(l′, l) =1 if l′ ∈ [l − α, l + α] and 0 otherwise, we can then use thefollowing objective function instead of (1).

max∑i∈Mj

z(m(i, j), L(p(i), p(j))) (3)

This leads us to find the p that maximizes the number of pathloss-matched edges in m and L. A simple example is given inFigures 2 and 3. Figure 2 shows the path loss measurementsof the users, and Figure 3 shows an RSS map generatedfrom a floor-plan by a simplified 3-ray trace model. Using(1) as an objective function, p = {userA : point3, userB :point2, userC : point1} minimizes the path loss matchingerror (i.e., |55−60|+ |90−85|). However, to relax sensitivityand find a larger set of feasible points, we can use (3).Using this objective function with the slack value α = 10,the path loss matching error obtained by positioning function{userA : point2, userB : point3, userC : point1} becomesequivalent to that obtained by another positioning function{userA : point3, userB : point2, userC : point1}. Thus,the results will contain a larger set of feasible points.

7

user A

user B

user C

55dB

90dB

Fig. 2. Example of pass loss measurements

point 1 point 2

point 3 point 4

80dB

85db 70dB

120dB

Fig. 3. Example of a RSS map generated from a floor-plan

4.2.2 Algorithm and ComplexityAssume a graph G = (Mj ,m) and a complete graph H =(N,N2). The above optimization problem is the maximumcommon subgraph isomorphism (MCSI) problem in graphtheory, which finds an induced graph G′ of G in H withthe objective function in (3). Although the general MCSIproblem is known to be NP-hard, our problem is a specialclass of MCSI in that we have a graph G with a star topologycentered at node j. Therefore, if j is allocated to pointv ∈ N , the calculation of the objective function in (3) canbe reduced in this case to the following problem: (i) for eachedge (i, j) ∈ m find all possible edges (u, v) ∈ N2 thatsatisfy z(m(i, j), L(u, v)) = 1 and (ii) find the positioningfunction p : Mj → N from (i) with the objective functionin (3). Part (i) needs to exhaustively test all edge pairs inG and H , respectively, and part (ii) can be reduced to themaximum bipartite matching problem on the graph, where Mj

and N are bipartite vertices, (j, v) are edges, and all possibleallocations of the nodes in Mj to points in N found in part(i) are also edges. The computation complexity of part (i) isO(|Mj ||N |), and that of part (ii) is also O(|Mj ||N |) by thepath matching algorithm [65]. Thus, to apply parts (i) and(ii) to all points in N , the complexity of our optimizationalgorithm is O(|Mj ||N |2). The pseudocode is given below.

Algorithm 1 Positioning algorithm for node j1: procedure find possible positioning functions (Mj , m, N , L)2: k ← 0; F = ∅3: for each v ∈ N do4: E ← (j, v)5: for each pair of edges (i, j) ∈ m and (u, v) ∈ N2 do6: E ← E ∪ (i, u) if z(m(i, j), L(u, v)) = 17: end for8: E′ ← bipartite graph matching result for ((Mj , N), E)9: if |E′| > k then

10: F ← E′; k ← |E′|11: else if |E′| == k then12: F ← F ∪ E′

13: end if14: end for15: return F as a set of positioning functions16: end procedure

The algorithm works as follows. In the loop between lines3 and 14, firstly, it is examined whether a pair of an edge fromnode j on the path loss measurement graph and an edge frompoint v on the simulated path loss graph satisfies function z

or not in the loop between lines 5 and 7. Then the bipartitematching algorithm is applied to set E of feasible edges in line8. Finally, the optimality of this result is examined comparedwith the best record in terms of the number of matched edgesbetween lines 9 and 13.

4.3 Location Convergence via Indoor Path Tracking

The mechanism for our indoor path tracking (i.e., deadreckoning) over a floor-plan uses commercial off-the-shelf(COTS) smartphone sensors to obtain movement distance andheading changes. The major challenge here is to use theunreliable inertial sensors in smartphones (e.g., magnetometersand accelerometers) to accurately track a user’s path.

We use Android’s magnetic field sensors and heading toidentify the direction of movement. During our initial im-plementation, we observed that the readings from a compasshave an internal bias in addition to the fluctuations caused byeven a slight sway, irregular motion, or by magnetic fieldsin the surroundings. We use a method recently proposedby Constandache et al. [66] to overcome above errors anddetect significant heading changes accurately at a corner whiletaking a turn and further use the following condition for moreaccurate turn detection:

Avg(t(i+1))−Avg(ti) =StdDev(ti) + StdDev(ti+1)

2(4)

where Avg(ti) denotes the average compass readings over ati time period, StdDev(ti) denotes the standard deviation ofcompass reading during ti, and G denotes a guard factor.

Next, we compute the distance traveled by a user as aproduct of the number of steps taken and step stride length.This is because, according to [66], the technique of double-integrating acceleration readings could induce a large erroreven over a small distance. Our algorithm continuously readsaccelerometer data to calculate user step count while it filtersout the noise. Then, it infers an increment in step count basedon changes in the observed readings.

We converted a sample floor-plan into an N × N matrix,as shown in Figure 4. Here, each coordinate point is 2 maway from its neighbors. This map comprises just the symbols{1, 0,W}, where “1” indicates accessible points, “0” indicatesinaccessible points, and “W” represents walls. After the initialWi-Fi scan, all possible 1s on this grid are listed. Every

8

0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0

wwwwwwwwwwww 1 1wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww 1 1 1wwwwwwwwwww

w 1 1 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1w 1 1 1 1w 1 1w 1 1w 1 1 1 1 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1www

wwwwwwwwwww 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1wwww 1w 1ww 1ww 1 1 1 1 1 1 1 1 1 1 1w 1 1 1wwwww 1 1 1www

w 1 1 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1www 1wwwwww 1w 1ww 1 1 1 1 1 1 1 1 1 1 1w 1 1 1wwwww 1 1 1www

w 1 1 1 1 1 1 1 1 1w 1 1 1 1wwwwwwwwwwwwwwwwww 1 1 1 1 1 1 1w 1 1 1 1 1w 1w 1w 1 1 1w 1 1w 1 1ww 1wwwwwwwwww 1 1 1wwwww 1 1 1www

w 1 1 1 1 1 1 1 1 1w 1 1 1 1wwwwwww 1w 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1w 1w 1 1 1w 1 1 1 1 1 1 1www 1 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1 1 1 1

wwwwwwwwwww 1 1 1 1wwwwwww 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1wwwww

w 1 1 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1 1 1w

wwwwwwwwwww 1 1 1 1w 1wwwww 1wwwwwwwwww 1 1 1 1 1 1 1wwwww 1w 1www 1wwwwwwwww 1wwwwwwww 1 1w 1 1 1wwwwwwwwwww

0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

0 0wwwwwwwww 1 1 1 1wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww 1 1 1wwwwwwwww 0 0

0 0w 1 1 1 1 1w 1w 1 1 1 1w 1w 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0

0 0wwwwww 1w 1 1 1 1 1w 1wwwwwww 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0wwwwwwwww 1 1 1wwwwwwwww 0 0

0 0w 1 1 1w 1 1 1w 1 1 1 1w 1w 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1ww 1w 1 1 1w 1w 1w 1 1 1w 0 0

0 0wwwwww 1ww 1 1 1 1w 1w 1wwwww 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0wwwwww 1 1w 1 1 1w 1w 1wwwww 0 0

0 0w 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1w 1 1w 1 1 1w 1 1 1 1 1 1 1w 0 0

0 0wwwwww 1 1w 1 1 1 1w 1 1wwwwww 0 0 0 0 0 0 0 0 0 0 0 0wwwwwwwwwwwww 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0wwwwww 1 1 1 1 1 1 1 1 1 1wwwww 0 0

0 0w 1 1 1w 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1w 0 0

0 0wwwwww 1 1w 1 1 1 1w 1 1wwwwww 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1w 0 0

0 0w 1 1 1w 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1w 0 0

0 0w 1 1 1w 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 1 1 1w 0 0

0 0wwwwwwwww 1 1 1 1wwwwwwwww 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0wwwwwwwww 1 1 1wwwwwwwww 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 0 0

0 0wwwwwwwww 1 1 1 1wwwwwwwww 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0wwwwwwwww 1 1 1wwwwwwwww 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1 1 1w 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0w 1 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1 1w 0 0

0 0wwwwwwwww 1 1 1 1wwwwwwwwwwwwwwwwwwwwwwww 1 1 1 1 1 1 1wwwwwwwwwwwwwwwwwwwwwwwwwww 1 1 1wwwww 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1w 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1w 1 1 1 1 1 1 1 1 1 1 1w 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1wwwwww 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1w 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1w 1 1 1 1w 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1wwwww 0 0 0 0 0 0

0 0wwwwwwwww 1 1 1 1wwww 1ww 1wwwwwwwwwwwwwwww 1 1 1 1 1 1 1wwww 1wwwww 1w 1wwwwwwwwwwwwww 1 1 1 1 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1w 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1ww 1wwwwww 1wwwww 1wwwwww 1wwwwwwwww 1wwwwww 1wwwww 1wwww 1 1wwwwwwwww 1ww 1wwwwww 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1w 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1w 1 1 1w 1 1 1 1 1 1w 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0

0 0w 1 1 1 1 1 1 1 1 1 1 1 1w 1 1 1 1 1 1 1wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww 1 1 1 1 1 1 1w 1 1 1 1 1 1 1w 0 0 0 0 0 0

0 0wwwwwwwwwwwwwwwwwwwwww 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0wwwwwwwwwwwwwwwww 0 0 0 0 0 0

Fig. 4. Computer Science department building map: 1(accessible), 0 (inaccessible), and W (wall) [53]

Fig. 5. Step Profiling App: generating a user’s stridelength at 3 mph

time a distance equal to the distance between two coordinateson the grid (2 m) is traveled by the user, the path trackingmodule records the user’s direction of movement. Based on theobserved movements, each coordinate in the list is updated,and coordinates that fall on a 0 or W are eliminated. Thisprocess of elimination continues iteratively as the user moveswhile changing direction, and eventually concretizes to asingle point on the grid that represents the user’s true position.Once one user locates himself/herself correctly, he/she canbroadcast his/her position to other team members so that theymay eliminate their invalid coordinates faster.

Accurate estimation of distance walked: Experimentally,we observed that walking speed also plays a crucial role inthe calculation of stride length. Therefore, we incorporatedwalking speed and the corresponding stride length into oursystem. Our profiling approach consists of two modes: profilecreation and profile usage.

Each user trains the system for his specific stride lengthsbased on his/her different walking speeds in the profile creationmode. The graphical user interface for this is shown in Figure5. It shows different speed options of 1 mph, 2 mph, and so onup to 7 mph. It also shows the corresponding lap time. For ourtests, we generated four separate profiles for users of differentgenders and heights. The test participants walked a test areaof length 100 m to train the system for different speeds, i.e.,step frequencies (by time-limiting the distance to be covered)

0

0.5

1

1.5

2

2.5

1 2 3 4 5 6 7

Stride length (m)

Step speed (mph)

Male, 5.90 feetMale, 5.74 feet

Female, 5.31 feetFemale, 5.02 feet

Fig. 6. Stride length as a function of walking speed (fordifferent genders and heights) [53]

and corresponding stride lengths.Figure 6 depicts our experiments results. A user’s stride

length increases, independently of height and gender, as thewalking speed increases. The stride length is also closelyrelated to the user’s walking style. For instance, a femaleof height 5.83 ft walking at a speed of 5 mph has a similarstride length as that of a male of height 5.90 ft. This exampleemphasizes why it is required to use step profiling for eachuser in order to obtain a more accurate traveled distance. Wenote that, according to the human kinematics literature [67],[68], it has been shown that human walking patterns are uniqueto individuals and do not vary significantly over time. Further,walking patterns can be well modeled based on walking speedand stride length, which provides a theoretical basis for ourwork.

TABLE 1Average stride length and step-profiling results for a male

(5.83 ft) at three different step speeds

Speed Mod. Fast RunDistance (m) 100 100 100Lap Time (s) 74.5 44.7 31.9

Observed # of Steps 127 77 68# of Steps (Ground Truth) 127 84 71

Male Stat. Step Len. (m) 0.737 0.737 0.737Dist Err. 6.4 42.04 49.884

Male Profile Step Len. (m) 0.79 1.28 1.42Dist Err. 0.33 1.44 3.44

In the profile usage mode, the system uses the existingprofile reference data to dynamically obtain a user’s stridelength from his/her current walking speed. Tables 1 and 2compare the error percentage (relative deviation from theground truth) in distance traveled by a male user and femaleuser, respectively, for average (statistical) stride length andprofiling stride length. According to [69] based on the averagestride length, the stride for a male of height 5.83 ft is asingle fixed value (0.737 m) for all walking speeds. Similarly,for a female of height 5.5 ft, it is 0.637 m. Because stridelength naturally becomes larger with walking speed, strideanalysis becomes more inaccurate at higher speeds, causingthe error percentage increase. However, profiling the stride

9

TABLE 2Average stride length and step-profiling results for female

(5.5 ft) at three different step speeds

Speed Mod. Fast RunDistance (m) 100 100 100Lap Time (s) 74.5 44.7 31.9

Observed # of Steps 121 77 75# of Steps (Ground Truth) 125 86 77

Female Stat. Step Len. (m) 0.673 0.673 0.673Dist Err. 18.93 48.179 49.525

Female Profile Step Len. (m) 0.82 1.32 1.42Dist Err. 0.78 1.64 5.08

Fig. 7. Preprocessed floorplan and Routes 1, 2, and 3 [53]

length is sufficient here. For example, for the stride lengthof a male of height 5.83 ft was observed at moderate, fast,and running speeds to be 0.79, 1.39, and 1.67 m, respectively.With profiling information, the difference between the groundtruth and calculated distance (as stridelength × stepcount)dropped significantly. This improved the system performancetremendously. A dynamic system is more suitable for ourtarget scenario, where a firefighter could be running andwalking intermittently.

To summarize, our profiling method consists of two modes,namely profile creation mode and usage mode. In the profilecreation mode, a user walks a test road for a given walkingspeed (i.e., step frequency), and the average stride length isthen estimated by dividing the road length by the step count.A user’s reference database is built by collecting stride lengthdata for different speeds. In the profile usage mode, the systemdynamically obtains a user’s stride length from the referencedatabase by simply determining the user’s current walkingspeed. Because walking patterns are unique to individuals anddo not vary significantly over time [67], [68], this manualcalibration can be performed once when a user initializes thesystem’s mobile client. Note that it is possible to automatethis calibration process, and thus, the mobile system canperiodically trigger calibration. For example, when a userwalks outside, it can record GPS tracks along with the walkingdata. We then can find a set of straight line segments fromthe GPS traces. The walking data collected while a user istraveling along the straight line segments can then be used toestimate stride length as we did in manual profiling.

5 EVALUATION

We evaluated CLIPS from an Android-based test bed toexperiment our system’s overall performance in real-worldscenarios, which is presented in Section 5.1. To verify thegeneral system performance of CLIPS in diverse locations(with floor-plans obtained from UCLA, Osaka University,and KAIST), we further performed extensive simulations andpresent the results in Section 5.2. Both experimental andsimulation results demonstrate that CLIPS lead to an accurateposition significantly faster than existing non-collaborativeschemes.

5.1 Test Bed Experiments5.1.1 Experiment SetupWe conducted our field tests in the Computer Science de-partment building at UCLA. As shown in Figure 7, we firstgenerated a preprocessed floor-plan with a grid overlay. Weused stride length profiles for accurate dead reckoning withvarying numbers of team members, ranging from two to nine.For path tracking (Figure 4), the same floor-plan was used.We deliberately chose three representative routes to furtherevaluate how the nature of a route traveled by a user (i.e.,the number of corner turns and the length of a straight linetaken by a user) can influence the system performance; theperformance results with random paths are reported in thesimulation section. As shown in Figure 7. Route 1, 2, and 3start from the same starting point, but while Route 1 containsa long straight path with just a single turn, Route 2 comprisesof more turns with long enough straight paths between twoconsecutive turns. Route 3 includes many turns with veryshort straight paths after each turn.

5.1.2 Experimental ResultsWe evaluate CLIPS’s performance with the followings: (1) theimpact of the team size and the slack value, (2) convergencespeed variation when taking different routes, (3) convergenceaccuracy with and without stride length profiles, and (4)overall convergence delay between CLIPS and smartphonedead reckoning.

Impact of team size and slack value: Figure 8 shows thefeasible coordinate ratio (FCR) for different numbers of peers.We also varied the slack value α to determine its optimumvalue. Note that we introduced α to cope with the differencebetween the floor-plan and reality, radio signal fluctuationover time, and Wi-Fi chipset disparity. Its value should besmall enough to eliminate most infeasible points, but largeenough to not miss the true location. The FCRs are calculatedby dividing the number of outcomes by all the coordinatesin the floor-plan (48 × 48 coordinates). While the α valuesshowed a positive relationship with the FCR, we observedthat the number of false positives decreases as we increasethe number of peers. Figure 9 illustrates the hit ratio, whichis the probability that the matching outcomes contain the truepositive position. As Figure 8 shows, we varied the number ofteam members as well as α to observe the changing systembehavior in different cases. The probability that a currentpoint is scanned becomes 100% as the α has been increased.

10

0

20

40

60

80

100

2 4 6 8 10 12 14Feasible Coordinates Ratio (%)

Alpha (dB)

1 mbr2 mbr3 mbr

4 mbr5 mbr6 mbr

7 mbr8 mbr

Fig. 8. Matching outcome ratio for different numbers ofteam members [53]

0

20

40

60

80

100

2 4 6 8 10 12 14

Hit Ratio (%)

Alpha (dB)

1 mbr2 mbr3 mbr4 mbr5 mbr6 mbr7 mbr8 mbr

Fig. 9. Probability of obtaining the current point fordifferent team sizes as a function of slack variable α [53]

Another interesting observation is that a smaller team (withfew members) requires a smaller α value to contain currentposition with the initial scan. For instance, a one-member teamwith an α value of 9 dB was observed to scan the currentposition with 100% probability, whereas an 8-member teamrequires an α value of 15 dB to obtain the same probability.

Convergence speed for different routes: of all the systemperformance metrics, accurate convergence speed to a uniquepoint is one of the most important for the proposed scheme. Weinvestigated how this factor is affected by different routes andby the utilization of peer-to-peer RSS exchanges and matchedit with simulation data. To verify this, we carefully chose threerepresentative routes: Route 1 (long straight paths and a turn),Route 2 (moderately long straight paths and a few turns),and Route 3 (short straight paths and many turns) to seewhich of the two factors, long straight paths or the numberof turns, affects the convergence speed more. We measured20 trials on each route and evaluated the success ratio of twodifferent stride approaches. Figure 10 shows the convergenceto the unique point as a function of the distance traveled.Route 1 shows a rapid drop in FCR in the beginning becauseof its distinct route feature, but after traveling 20 m, theroute showed relatively slow convergence. This was becausemultiple points on the straight path still remain as candidates.However, when a turn is taken after 60 m, it converges to aunique point. Route 2 shows the slowest convergence speed.This is because this route contains relatively moderate events,such as moderately long straight paths and turns. Route 3shows the fastest convergence speed to a unique point. Thisis because this route contains many turns. Thus, based on ourobservation, a complicated route tends to expedite the processof convergence to a unique point.

Convergence accuracy with stride length profiling: accu-rate dead reckoning with stride length profiling is CLIPS’sone of import features. In order to to evaluate this feature,two different stride length approaches, namely average stridelength (fixed) and stride length profiling (variable), have beentested 20 times per route. Given that the users are navigatingindoors, it is highly necessary to have accurate displacement

logging, as in the map shown in Figure 4. Our step profilinguses dynamic stride length to calculate the traveled distanceunlike the average stride length approach. Eliminating falsepositives on the floor-plan using this scheme is more accurate.For instance, at the end of a corridor, changes in directionthat are both earlier and later than required will cause theconvergence to a unique position to fail. Figure 11 shows theconvergence ratio of the two different mechanisms (fixed vs.variable), which are well matched with distance deviationsfor both the male and female users shown in Tables 1 and2, respectively. Step profiling provides 100% accuracy forRoute 1 while slightly degraded performance for Route 3 isobserved. This can be attributed to the error in compass read-ings caused by the magnetic fields in the surroundings. Evenwhen profiling correctly calculates the distance traveled, theerror in orientation can cause inaccurate results. To eliminatethese errors, a gyroscope and an accelerometer can be used inaddition to the orientation sensor, as explained in [48]. We willincorporate this feature in the future. In our experiments, theaverage stride length approach significantly underestimated theactual distance traveled in a long, straight corridor, resulting inpoor performance (32%.) However, for a path such as that ofRoute 3, the average stride length showed relatively reliableperformance. This could be because the overall distance erroris refreshed after each turn on this route.

Overall convergence delay of CLIPS: the aggregated timetaken to lead to a unique point with step profiling for eachof the three routes is shown in Figure 12. The delay dif-ference mainly occurs because users are traveling differentroutes while the initial Wi-Fi scanning and matching tookalmost constant time for all three scenarios. Overall delay isdominated by the traveling time while the Wi-Fi scanning andmatching delay are considerably smaller comparatively.

Smartphone dead reckoning vs. CLIPS: to evaluate howRSS matching can affect the overall system performance, wecompared dead reckoning using COTS smartphone sensorswith CLIPS. Figure 13 shows the exact distances that had tobe traveled by each system to converge to a unique locationon each of the three routes. We observed that for all three

11

0

5

10

15

20

25

30

20 40 60 80 100Feasible coordinate ratio (%)

Trveled distance (m)

Route 2

Route 1

Route 3

Fig. 10. Convergence speed of different routes [53]

0

10

20

30

40

50

60

70

80

90

100

Route 1

Route 2

Route 3

Convergence ratio (%)

statprofile

Fig. 11. Convergence success ratio of twostride length estimation methods for three differentroutes [53]

routes, smartphone dead reckoning required the user to travela longer distance in order to converge to a unique point. Figure14 shows that more detailed, feasible coordinates are lost asa function of the distance traveled. We note that smartphonedead reckoning shows that Route 2 requires a longer distanceto converge to a unique point, which consequently takes thelongest time compared to the other two routes. This is becausesmartphone dead reckoning shows more dependency on routes,as one route contains more turns to expedite convergenceprocess to the unique point. However, CLIPS starts from adifferent set of feasible points; therefore, it is less dependenton turns. Considering the results of Figures 12 and 13,we conclude that Wi-Fi scanning and matching improve thesystem performance with relatively little cost.

5.2 Simulations

5.2.1 Simulation Setup

To test the general system performance at different locations,we further evaluated the two major components of CLIPS,namely, matching RSS with path loss estimation and deadreckoning via network simulations using Qualnet. For thesimulations, floor-plans were obtained from UCLA, OsakaUniversity, and KAIST. To generate path loss estimation ona floor-plan, the wireless signal propagation was computedby Wireless Insight with a 3D model of the floor-plan (byCoCreate Modeling Software [70]). We used N × M gridtopologies, in which the side length of each grid is 2 m. Themeasured RSS was randomly generated from the simulatedRSS with Gaussian error N(0 dB, 5 dB) (mean 0 and variance= 5 dB). To verify the sensitivity of the matching betweensimulated and measured RSS, we set the default value of αto 10 dB; we also assumed that dead reckoning is accurate.In our simulations, we adopted random room search mobility(RRSM) to emulate a team member’s mobility. Initially, eachnode is located at a randomly chosen position. Each nodechooses a destination room randomly and moves to the desti-nation through a passage. When it reaches the destined room,it again randomly picks and moves to the next destination

room. The nodal speed was set to 2.5 m/s. Unless otherwisespecified, we report an average value of 50 runs with RRSMto mitigate any dependence on specific route paths and tounderstand general system behavior.

5.3 Simulation ResultsEffect of slack variable α on feasible coordinate and hit ratios:Figures 15 and 16 examine the effects of the slack variableα on the feasible coordinate ratio (FCR) and the hit ratio,respectively. To verify these, we varied the number of deployednodes from 1 to 30, and changed the slack variable from 2 to10. Figure 15 shows that all coordinates are considered asfeasible in the single node case. This is because the nodecan be deployed in any place on the passage and cannotverify feasibility with itself. As we increased the number ofdeployed nodes, the FCR decreased; the greater number ofnodes deployed, the smaller the set of feasible matches onthe map. Note that the number of coordinates increases aswe increase α. Figure 16 also examines the effect of α onthe hit ratio. This figure represents the performance accuracyof CLIPS based on the slack variable, coupled with Figure15. The greater the number of coordinates considered feasiblethanks to the larger α value, the higher the hit ratio. Byperforming this simulation, we confirmed that the relationshipbetween the number of deployed nodes and FCR as well asbetween the slack variable and hit ratio are well matched withthe test bed results.

Effect of team size on convergence speed: Figure 17 reportshow different numbers of team members can affect the conver-gence speed. All eight cases, from 1- to 15-node teams, showcorrelated behavior: FCR decreases as the traveled distanceincreases. We also note that teams with a large number ofmembers start with a smaller FCR, which means that a largeteam will require less traveled distance (and time) to convergeto unique coordinates.

Positioning Errors: Table 3 shows the average positioningerrors and their standard deviations. Positioning errors arecalculated as the average distance error from the feasiblecoordinates to true coordinates. CLIPS (with RSS) has a

12

0

20

40

60

80

100

120

140

160

Route 1

Route 2

Route 3

Cum

ula

tive

Tim

e (s

ec)

WiFi ScanningMatchingMoving

Fig. 12. Cumulative latencies of three modules fordifferent routes [53]

0

10

20

30

40

50

60

70

80

90

100

110

Route 1

Route 2

Route 3

Traveld distance (m)

Dead ReckoningCLIPS

Fig. 13. Distance that must be traveled to convergeto a unique point on three different routes: CLIPS vs.smartphone dead reckoning [53]

0

20

40

60

80

100

20 40 60 80 100Feasible coordinate ratio (%)

Trveled distance (m)

Route 3: Dead ReckoningRoute 1: Dead ReckoningRoute 2: Dead Reckoning

Route 2: CLIPSRoute 1: CLIPSRoute 3: CLIPS

Fig. 14. Convergence speed: CLIPS vs. smartphonedead reckoning

2

4

6

8

10

0

20

40

60

80

100

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Slack

Variable(dB)

FeasibleCoordinate

Ratio(%

)

Number of Nodes

80 100

60 80

40 60

20 40

0 20

Fig. 15. Effect of slack variable α on feasible coordi-nates

clear advantage over the scheme without RSS in terms of theaverage errors, but it has larger standard deviations. However,these deviations converge and become smaller after at least 50m has been traveled. The positioning error of CLIPS is finallyreduced to 2.2 m with σ = 1.4 m, which is reasonably small.

Sharing location updates: After a user has determined thecurrent location, periodically sharing this information withfellow team members (via ad hoc communications or a cen-tralized server) helps the team to eliminate false coordinates.To verify the improvement in performance caused by thislocation sharing, we evaluated the FCR based on the numberof determined locations. In Figure 18, we present the feasiblecoordinates of users as a function of the number of userswith known coordinates (acting as mobile anchors). The graphshows that the feasible coordinates exponentially decrease withthe number of users with known coordinates. Even for a singlenode, the FCR is reduced from 51% to 13%, and with anadditional node, it is further reduced to 2.3%. This result showshow much the overall system performance can be improvedby sharing determined locations.

6 DISCUSSION AND FUTURE WORK

Aggressive information sharing: In our simulation results, weshow that when nodes share discovered location information(act as mobile anchors), this can significantly reduce the sizeof feasible coordinates. Another way of reducing the sizeof feasible coordinates is to share the current progress oflocalization with other members (i.e., sharing the feasiblecoordinates of a node with the rest of the nodes). If nodeA receives the feasible coordinate of another node B, node Acan further eliminate the infeasible coordinates of node B inits feasible coordinates. We leave this simple improvement aspart of our future work.

Robust dissemination: For beaconing and location sharing,we use best effort delivery with user datagram protocol (UDP)broadcasting. Wireless errors can be effectively handled withperiodic packet broadcasting. In some cases, location updatesmight need to be delivered over multiple hop. By piggybackingdiscovered location information in a packet, the locationinformation of group members can be eventually disseminatedto the entire group.

Exploiting environmental signatures: Mobile devices with

13

2

4

6

8

10

0

10

20

30

40

50

60

70

80

90

100

1 4 7 10 13 16 19 22 25 28 Slack

Variable(dB)

HitRatio(%

)

Number of Nodes

90 100

80 90

70 80

60 70

50 60

40 50

30 40

20 30

10 20

0 10

Fig. 16. Effect of slack variable α on hit ratio

0

10

20

30

40

50

60

70

80

90

100

2 6 10 14 18 22 26 30 34 38 42 46 50 54 58 62 66 70

FeasibleCoordinate

Ratio(%

)

Travelled Distance (m)

mbr01

mbr03

mbr05

mbr07

mbr09

mbr11

mbr13

mbr15

Fig. 17. Effect of team size on convergence speed

TABLE 3Positioning Errors (m)

Traveled Distance (m) 10 20 30 40 50 60 70 80w/ RSS Average 19.85 17.01 14.31 10.55 6.82 4.14 3.31 2.21

Std Dev. 6.04 6.61 6.54 5.73 4.32 2.54 1.85 1.40w/o RSS Average 40.98 36.42 28.36 24.95 20.84 12.35 8.70 4.55

Std Dev. 1.05 1.09 1.29 2.22 1.70 3.07 3.28 1.24

0

10

20

30

40

50

60

70

0 5 10 15

FeasibleCoordinate

Ratio(%

)

Number of Known Coordinates (out of 15 nodes)

Feasible Coordinate Ratio

Fig. 18. Effect of known coordinate landmarks onconvergence speed and accuracy

their sensor measurements are dead reckoned while,in orderto detect unique environmental signatures within the build-ing (e.g., Wi-Fi RSS drops and magnetic fields), the samemeasurements are leveraged. Similarly to the recent workby Wang et al. [19], these unique signatures can be usedto fix errors caused by the dead reckoning, which in turnimproves the location accuracy of these signatures. However,surveying such environmental signatures takes considerabletime. One possible solution is to use an ad hoc deployment ofmobile beacons during the emergency operations to create “adhoc environmental signatures,” which is an idea inspired by

iBeacon scenarios. When a team of first responders arrivesat the floor of interest, they could scatter mobile beaconsthroughout the site. Then, CLIPS can leverage both mobileand stationary beacons. As shown in our simulation results,we can clearly see that as the number of beacons increases,we can significantly reduce the feasible coordinates, and shar-ing update information will further improve the convergencespeed. We leave the examination this hybrid scenario as a partof our future work.

Path-loss simulation of random coordinates: A random in-stead of uniform distribution of coordinates could be beneficialto remove flip and rotation issues in matching. The density ofsuch points is also an important parameter for understandingthe trade-off between accuracy and computation complexity.It might be possible to aggregate two points that have similarpropagation characteristics (i.e., points in the same smallroom) to reduce complexity and expedite convergence to theunique point. To do this automatically, we need to verifywhether two nearby points match after mapping.

Simultaneous multi-floor operations: If there are other re-sponders located on different floors, their Wi-Fi signals mayinterfere with those located on the current floor. In thisscenario, we can use the following strategies. One approachis to remove nodes whose RSS values are lower than somethreshold value. This is based on the fact that if mobile usersare located on different floors, their RSS values could besignificantly lower than those of the others. In addition, themobility of users can be considered: i.e., it is highly unlikelythat a user’s RSS trace measured on a given floor matcheswith that on a different floor. Note that we can further extendour algorithm by simultaneously considering multiple floorsand quickly determining which floor map is the best candidate.

14

Another approach is to use an extra sensor, namely a barometer(or altimeter), which is widely available in recent smartphones(e.g., Google Nexus 5). According to Muralidharan et al. [71],we can accurately detect whether users are located on the samefloor by whether their atmospheric pressure values are withinsome threshold range (e.g., for every 10 m in height, there isa pressure difference of 1.2 hPa). Furthermore, when a usermoves across floors, the change of pressure value is significantenough to detect this event.

7 CONCLUSIONS

In this paper, we consider time-critical indoor scenarios suchas search and rescue missions in a large building wheretraditional infrastructure-based localization schemes are notfeasible. For this scenario, we developed CLIPS, a novelinfrastructure-free, collaborative localization system. Unlikeinfrastructure-based approaches, CLIPS leverages peer-to-peerbeaconing and floor maps, which are readily available fortime-critical team operations. In CLIPS, mobile nodes initiallyestimate candidate coordinates in the floor maps by exchang-ing beacon signals and signal space maps. Mobility of usersis then tracked with dead reckoning, and this information isused to validate whether a movement on the map is feasiblewhen those candidate coordinates were considered as startingpositions. This approach can easily remove all the invalidcoordinates, thereby quickly leading to position fixes. Weevaluated the proposed system in both Android-based testbedexperiments and extensive simulations. Our results confirmedthat CLIPS can provide accurate localization performance andachieve much lower delay for position fixes.

ACKNOWLEDGEMENT

This work was supported in part by JSPS KAKENHIJP15H02690 and JP15K12019. This research was also sup-ported in part by Basic Science Research Program throughthe National Research Foundation of Korea(NRF) fundedby the Ministry of Education (NRF-2016R1C1B2011415),Inha University Research Grant (54465-01), and Institute forInformation & communications Technology Promotion (IITP)grant funded by the Korea government (MSIP) (No. B0101-15-1272, Development of Device Collaborative Giga-Level SmartCloudlet Technology).

REFERENCES

[1] H. Shin, Y. Chon, K. Park, and H. Cha, “FindingMiMo: tracing a missingmobile phone using daily observations,” in MobiSys, 2011.

[2] K. Chintalapudi, A. Padmanabha Iyer, and V. N. Padmanabhan, “Indoorlocalization without the pain,” in MobiCom, 2010.

[3] “Skyhook, XPS, Web Site,” http://www.skyhookwireless.com.[4] “Apple, MobileMe, Web Site,” http://www.me.com.[5] “Google, Latitude, Web Site,” https://www.google.com/latitude.[6] P. Bahl and V. Padmanabhan, “RADAR: an in-building RF-based user

location and tracking system,” in Infocom, 2000.[7] I. Smith, J. Tabert, T. Wild, A. Lamarca, A. Lamarca, Y. Chawathe,

Y. Chawathe, S. Consolvo, S. Consolvo, J. Hightower, J. Hightower,J. Scott, J. Scott, T. Sohn, T. Sohn, J. Howard, J. Howard, J. Hughes,J. Hughes, F. Potter, F. Potter, P. Powledge, P. Powledge, G. Borriello,G. Borriello, B. Schilit, and B. Schilit, “Place lab: Device positioningusing radio beacons in the wild,” in Pervasive, 2005.

[8] G. Borriello, A. Liu, T. Offer, C. Palistrant, and R. Sharp, “WALRUS:wireless acoustic location with room-level resolution using ultrasound,”in MobiSys, 2005.

[9] S. P. Tarzia, P. A. Dinda, R. P. Dick, and G. Memik, “Indoor localizationwithout infrastructure using the acoustic background spectrum,” inMobiSys, 2011.

[10] R. Want, A. Hopper, V. Falcao, and J. Gibbons, “The active badgelocation system,” ACM Trans. Inf. Syst., 1992.

[11] J. Guerrieri, M. Francis, P. Wilson, T. Kos, L. Miller, N. Bryner,D. Stroup, and L. Klein-Berndt, “RFID-assisted indoor localization andcommunication for first responders,” in EuCAP, 2006.

[12] G. yao Jin, X. yi Lu, and M.-S. Park, “An indoor localization mechanismusing active RFID tag,” in SUTC, 2006.

[13] L. Ni, Y. Liu, Y. C. Lau, and A. Patil, “LANDMARC: indoor locationsensing using active RFID,” in PerCom, 2003.

[14] J. Leonard and H. Durrant-Whyte, “Simultaneous map building andlocalization for an autonomous mobile robot,” in IROS, 1991.

[15] B. Ferris, D. Fox, and N. Lawrence, “WiFi-SLAM using Gaussianprocess latent variable models,” in IJCAI, 2007.

[16] H. Liu, Y. Gan, J. Yang, S. Sidhom, Y. Wang, Y. Chen, and F. Ye, “Pushthe limit of WiFi based localization for smartphones,” in Mobicom, 2012.

[17] Y. Kim, Y. Chon, and H. Cha, “Smartphone-based collaborative andautonomous radio fingerprinting,” Trans. Sys. Man Cyber Part C, vol. 42,no. 1, pp. 112–122, Jan. 2012.

[18] H. Shin, Y. Chon, and H. Cha, “Unsupervised construction of an indoorfloor plan using a smartphone,” Systems, Man, and Cybernetics, PartC: Applications and Reviews, IEEE Transactions on, vol. 42, no. 6, pp.889–898, Nov 2012.

[19] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R.Choudhury, “No need to war-drive: Unsupervised indoor localization,”in MobiSys, 2012.

[20] C. Luo, H. Hong, and M. C. Chan, “Piloc: A self-calibrating participa-tory indoor localization system,” in IPSN, 2014.

[21] C. Wu, Z. Yang, and Y. Liu, “Smartphones based crowdsourcing forindoor localization,” Mobile Computing, IEEE Transactions on, vol. 14,no. 2, pp. 444–457, Feb 2015.

[22] N. B. Priyantha, A. Chakraborty, and H. Balakrishnan, “The Cricketlocation-support system,” in MobiCom, 2000.

[23] A. Ward, A. Jones, and A. Hopper, “A new location technique for theactive office,” Personal Communications, IEEE, vol. 4, no. 5, pp. 42–47,Oct 1997.

[24] J. Xiong, K. Sundaresan, and K. Jamieson, “Tonetrack: Leveragingfrequency-agile radios for time-based indoor wireless localization,” inMobiCom, 2015.

[25] M. Youssef and A. Agrawala, “The Horus WLAN location determinationsystem,” in MobiSys, 2005.

[26] A. Varshavsky, E. de Lara, J. Hightower, A. LaMarca, and V. Otsason,“Gsm indoor localization,” Pervasive Mob. Comput., vol. 3, no. 6, pp.698–720, Dec. 2007.

[27] L. Li, G. Shen, C. Zhao, T. Moscibroda, J.-H. Lin, and F. Zhao, “Ex-periencing and handling the diversity in data density and environmentallocality in an indoor positioning service,” in MobiCom, 2014.

[28] “Koozyt, Inc., PlaceEngine, Web Site,”http://www.placeengine.com/showe/about.

[29] M. Azizyan, I. Constandache, and R. R. Choudhury, “SurroundSense:mobile phone localization via ambience fingerprinting,” in MobiCom,2009.

[30] R. Gao, Y. Tian, F. Ye, G. Luo, K. Bian, Y. Wang, T. Wang, and X. Li,“Sextant: Towards ubiquitous indoor localization service by photo-takingof the environment,” Mobile Computing, IEEE Transactions on, vol. PP,no. 99, pp. 1–1, 2015.

[31] Y. Gu, A. Lo, and I. Niemegeers, “A survey of indoor positioningsystems for wireless personal networks,” Communications Surveys Tu-torials, IEEE, vol. 11, no. 1, pp. 13–32, First 2009.

[32] H. Liu, H. Darabi, P. Banerjee, and J. Liu, “Survey of wireless indoorpositioning techniques and systems,” Systems, Man, and Cybernetics,Part C: Applications and Reviews, IEEE Transactions on, vol. 37, no. 6,pp. 1067–1080, Nov 2007.

[33] H. Lim, L.-C. Kung, J. C. Hou, and H. Luo, “Zero-configuration indoorlocalization over IEEE 802.11 wireless infrastructure,” Wirel. Netw.,vol. 16, no. 2, pp. 405–420, Feb. 2010.

[34] Y. Gwon and R. Jain, “Error characteristics and calibration-free tech-niques for wireless LAN-based location estimation,” in MobiWac, 2004.

[35] H. Lim, L.-C. Kung, J. C. Hou, and H. Luo, “Zero-Configuration, RobustIndoor Localization: Theory and Experimentation,” in Infocom, 2006.

[36] Y. Ji, S. Biaz, S. Pandey, and P. Agrawal, “ARIADNE: a dynamic indoorsignal map construction and localization system,” in MobiSys, 2006.

15

[37] S. Sen, J. Lee, K.-H. Kim, and P. Congdon, “Avoiding multipath torevive inbuilding wifi localization,” in MobiSys, 2013.

[38] A. T. Mariakakis, S. Sen, J. Lee, and K.-H. Kim, “Sail: Single accesspoint-based indoor localization,” in MobiSys, 2014.

[39] L. Y. Suining He, S.-H. Gary Chan and N. Liu, “Fusing noisy finger-prints with distance boundsfor indoor localization,” in Infocom, 2015.

[40] C. X. C. Y. Y. L. Chenshu Wu, Zheng Yang and M. Liu, “Staticpower of mobile devices: Self-updating radio maps for wireless indoorlocalization,” in Infocom, 2015.

[41] L. E. Holmquist, J. Falk, and J. Wigstrom, “Supporting group collabo-ration with interpersonal awareness devices.” Personal and UbiquitousComputing, vol. 3, no. 1/2, pp. 13–21, 1999.

[42] J. Krumm and K. Hinckley, “The nearme wireless proximity server,” inUbiComp, 2004.

[43] “Bluehoo, Web Site,” http://bluehoo.com.[44] K. A. Li, T. Y. Sohn, S. Huang, and W. G. Griswold, “Peopletones: A

system for the detection and notification of buddy proximity on mobilephones,” in MobiSys, 2008.

[45] C. Peng, G. Shen, Y. Zhang, Y. Li, and K. Tan, “Beepbeep: A highaccuracy acoustic ranging system using COTS mobile devices,” inSenSys, 2007.

[46] N. Banerjee, S. Agarwal, P. Bahl, R. Ch, A. Wolman, and M. Corner,“Virtual compass: Relative positioning to sense mobile social interac-tions,” in Pervasive, 2010.

[47] T. Higuchi, S. Fujii, H. Yamaguchi, and T. Higashino, “Mobile nodelocalization focusing on stop-and-go behavior of indoor pedestrians,”Mobile Computing, IEEE Transactions on, vol. 13, no. 7, pp. 1564–1578, July 2014.

[48] O. Woodman and R. Harle, “Pedestrian localisation for indoor environ-ments,” in UbiComp, 2008.

[49] K. chan Lan and W.-Y. Shih, “Using smart-phones and floor plans forindoor location tracking,” Human-Machine Systems, IEEE Transactionson, vol. 44, no. 2, pp. 211–221, April 2014.

[50] D. Philipp, P. Baier, C. Dibak, F. Durr, K. Rothermel, S. Becker,M. Peter, and D. Fritsch, “Mapgenie: Grammar-enhanced indoor mapconstruction from crowd-sourced data,” in Pervasive Computing andCommunications (PerCom), 2014 IEEE International Conference on,March 2014, pp. 139–147.

[51] S. Sorour, Y. Lostanlen, S. Valaee, and K. Majeed, “Joint indoorlocalization and radio map construction with limited deployment load,”Mobile Computing, IEEE Transactions on, vol. 14, no. 5, pp. 1031–1043,May 2015.

[52] E. Martin, O. Vinyals, G. Friedland, and R. Bajcsy, “Precise indoorlocalization using smart phones,” in MM, 2010.

[53] Y. Noh, H. Yamaguchi, U. Lee, P. Vij, J. Joy, and M. Gerla, “Clips:Infrastructure-free collaborative indoor positioning scheme for time-critical team operations,” in PerCom, March 2013, pp. 172–178.

[54] P. Dosch, K. Tombre, C. Ah-Soon, and G. Masini, “A complete systemfor the analysis of architectural drawings,” IJDAR, vol. 3, no. 2, pp.102–116, 2000.

[55] “Google Maps Floor Plans, Web Site,”http://maps.google.com/help/maps/floorplans/.

[56] H. Ye, T. Gu, X. Zhu, J. Xu, X. Tao, J. Lu, and N. Jin, “Ftrack:Infrastructure-free floor localization via mobile phone sensing,” inPerCom, 2012.

[57] C. R. Anderson and T. S. Rappaport, “In-building wideband partitionloss measurements at 2.5 and 60 ghz.” IEEE Transactions on WirelessCommunications, vol. 3, no. 3, pp. 922–928, 2004.

[58] H. Kim and H. Ling, “Electromagnetic scattering from an inhomoge-neous object by ray tracing,” Antennas and Propagation, IEEE Trans-actions on, vol. 40, no. 5, pp. 517–525, May 1992.

[59] R. Valenzuela, “Ray tracing prediction of indoor radio propagation,” inPIMRC, 1994.

[60] A. Falsafi, K. Pahlavan, and G. Yang, “Transmission techniques for radioLAN’s - A comparative performance evaluation using ray tracing,” IEEEJournal on Selected Areas in Communications, vol. 14, no. 3, pp. 477–491, 1996.

[61] M. Lott, “On the Performance of an Advanced 3D Ray Tracing Method,”in EW, 1999.

[62] K. A. Remley, H. R. Anderson, and A. Weisshaar, “Improving theaccuracy of ray-tracing techniques for indoor propagation modeling.”IEEE T. Vehicular Technology, vol. 49, no. 6, pp. 2350–2358, 2000.

[63] “Insite, Web Site,” http://www.remcom.com/wireless-insite.[64] ITU-R, “Propagation data and prediction methods for the planning of

indoor radiocommunication systems and radio local area networks in thefrequency range 300MHz to 100GHz,” Recommendation ITU-R P.1238-8, 2015.

[65] T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson, Introductionto Algorithms, 3rd ed. McGraw-Hill Higher Education, 2001.

[66] I. Constandache, X. Bao, M. Azizyan, and R. R. Choudhury, “Did yousee Bob?: human localization using mobile phones,” in MobiCom, 2010.

[67] R. Margaria, Biomechanics and energetics of muscular exercise. Claren-don Press, 1976.

[68] A. Zijlstra, E. de Bruin, N. Bruins, and W. Zijlstra, The Step Length-frequency Relationship in Physically Active Community-dwelling OlderWomen. Springer, 2008.

[69] “Step size in pedometers, Web Site,”http://walking.about.com/cs/pedometers/a/pedometerset.htm.

[70] “Parametric Technology Corporation, Web Site,”http://www.ptc.com/products/creo-elements-direct/modeling-express/.

[71] K. Muralidharan, A. J. Khan, A. Misra, R. K. Balan, and S. Agarwal,“Barometric phone sensors: more hype than hope!” in HotMobile, 2014.

Dr. YoungTae Noh is an assistant professorin the Department of Computer Science andInformation Engineering, Inha University. Hereceived his B.S. in computer science fromChosun University in 2005, an M.S. degree inInformation and Communication from GwangjuInstitute of Science Technology (GIST) in 2007,and a Ph.D. in computer science at University ofCalifornia, Los Angeles (UCLA) in 2012. Beforejoining Inha University, he was a staff memberat Cisco Systems and a postdoc research as-

sociate at Purdue University. His research areas include data centernetworking, wireless networking, and mobile computing.

Dr. Hirozumi Yamaguchi is currently an asso-ciate professor at Osaka University. He receivedthe BE, ME, and PhD degrees in informationand computer sciences from Osaka University,Japan, in 1994, 1996, and 1998, respectively.His current research interests include designand development of mobile and pervasive sys-tems. He is a member of the IEEE.

Dr. Uichin Lee is an associate professor inthe Department of Knowledge Service Engi-neering at Korea Advanced Institute of Scienceand Technology (KAIST). He received a B.S.in computer engineering from Chonbuk NationalUniversity in 2001, an M.S. degree in computerscience from KAIST in 2003, and a Ph.D. degreein computer science from the University of Cal-ifornia at Los Angeles (UCLA) in 2008. Beforejoining KAIST, he was a member of technicalstaff at Bell Laboratories, Alcatel-Lucent until

2010. His research interests include social computing systems andmobile/pervasive computing.