Communicating Is Crowdsourcing: Wi-Fi Indoor Localization ...

Zhao JZ, Jiang ZP, Li XY et al.. JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY : 1– May 2014

Communicating Is Crowdsourcing: Wi-Fi Indoor Localization with CSI-based Speed Estimation

Member CCF, ACM, IEEE, Ji-Zhong Zhao(赵季中)1, Zhi-Ping Jiang(蒋志平)1, Member ACM, SeniorMember IEEE, Xiang-Yang Li(李向阳)2,3, Member IEEE, Shao-Jie Tang(唐少杰)4, Member CCF, ACM,IEEE, Jin-Song Han(韩劲松)1, Member CCF, ACM, IEEE, Wei Xi(惠维)1, Kun Zhao(赵鲲)1, and ZhiWang(王志)1

1School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an, P.R.C2Department of Computer Science and Technology and TNLIST, Tsinghua University3Department of Computer Science, Illinois Institute of Technology, IL, U.S4Department of Information Science, UT Dallas, TX, U.S

E-mail: {zjz,hanjinsong,zhaokun2012,zhiwang}@mail.xjtu.edu.cn; {jiangzp.cs, weixi.cs, tangshaojie,lirui.xjtu}@gmail.com ; [email protected]

Received submit date

摘要 Numerous indoor localization techniques have been proposed recently to meet the intensive demand forlocation based service (LBS). Among them, Wi-Fi fingerprint-based approaches are the most popular solutions,and the core challenge of that is to lower the cost of fingerprint site-survey. One of the trends is to collect thepiecewise data from clients and establish the radio map in crowdsourcing manner, however the low participationrate blocks the practical use.

In this work, we propose a passive crowdsourcing CSI-based indoor localization scheme, C2IL. Despite acrowdsourcing based approach, our scheme is totally transparent to client except the only requirement is to connectto our 802.11n APs. C2IL is built upon an innovative method to accurately estimate the moving speed solelybased on 802.11n Channel State Information (CSI). Knowing the walking speed of a client and its surroundingAPs, a graph matching algorithm is employed to extract the RSS fingerprints and establish the fingerprint map.In localization phase, we design a trajectory clustering-based localization algorithm to provide precise real- timeindoor localization and tracking. We developed and deployed a practical working system of C2IL in a large officeenvironment. Extensive evaluations indicate that the error of speed estimation is within 3%, and the localizationerror is within 2m at 80% time in very complex indoor environment.

Keywords CSI, Speed Estimation, Fingerprint Map, Graph Matching, Indoor Localization

This work is supported by the National Natural Science Foundation of China (NSFC) under Grant Nos. 61325013,61190112, 61170216, and 61228202; Natural Science Foundation of U.S (NSF) under Grant Nos. CNS-0832120, CNS-1035894, ECCS-1247944, ECCS-1343306; the Fundamental Research Funds for the Central Universities of China underProject No. 2012jdgz02 (Xi’an Jiaotong University), and the Research Fund for the Doctoral Program of Higher Educationunder project No. 20130201120016.

2 J. Comput. Sci. & Technol., May. 2014, ,

1 Introduction

With the prosperity of mobile devices, espe-cially smartphones, location based services (LBS),which use the geographic position to provide tar-geted services, have become pervasive to provideadded value of existing services. A critical chal-lenge of LBS is to find the accurate location ofmobile devices. GPS has successfully dominatedthe outdoor localization. Unfortunately in indoorenvironment, the most facile wireless received sig-nal strength (RSS) is neither accurate nor consis-tent due to the highly dynamic and complex en-vironment. As the flourishing of smartphones andcrowdsourcing computation models, numerous in-door localization techniques have been proposed tocollect fractional environment features and collab-oratively provide precise indoor localization.

The relatively good accuracy and simplicityof fingerprint based localization schemes has at-tracted massive of effort in the community. Wi-Fi fingerprint- based schemes can provide meter-level indoor localization accuracy at the expenseof explicit site-survey. Its high deployment costand low adaptiveness to environment change hin-ders the practical effectiveness. Recently severalnovel techniques, e.g., [1–5], have been proposedto raise the usability and accuracy. Among theseapproaches, a hot research trend is to incorpo-rate crowdsourcing model and built-in sensors intoday’s smartphone. LiFS [5] reduces the site-survey by using the moving distance, estimatedfrom counting the number of steps by accelerome-ter, as constraints for matching between the mapand trace-graph, achieving average accuracy of5.8m Zee [2] achieves a mean accuracy 3m byestimating the moving direction and moving dis-tance by similarly leveraging the sensors in smart-phone. Centuar [3] and PAL [4] both calibrate the

Wi-Fi fingerprints database using acoustic rang-ing and they achieve 1∼3m accuracy. All theseapproaches significantly improved the practicabil-ity of Wi-Fi based indoor localization, however, webelieve there are plenty of room for improving thelocalization accuracy while reducing or even elimi-nating the dependence of site-survey and noisy in-ertial sensors.

In this work, we design and develop an indoorlocalization system C2IL with even lower cost andhindrance. Our scheme exploits the Channel StateInformation (CSI) introduced in 802.11n Spec., forbetter speed estimation which in turn results in abetter fingerprints matching and localization. Ad-ditionally, C2IL, as in most recent techniques, col-lects the WiFi RSS fingerprints during communi-cation between the user and WiFi APs, and pop-ulates the fingerprints database by crowdsourcing.C2IL has good performance in complex environ-ment with rich multipath effect, while several re-cently developed schemes suffer from lower local-ization accuracy in such a complex environment.

The accurate moving speed estimation isbased on the readily available CSI in IEEE 802.11nsystem that uses OFDM and Multi-Input & Multi-Output (MIMO) technologies. CSI measures theinherent channel propagation decay for each sub-carrier of the underlying OFDM system. Since CSIfocuses more on small- scale wireless fading, it hasadvantages on capturing the mobile channel char-acteristics. In our preliminary testes carried in dif-ferent scenarios, we found the CSI amplitude ex-hibits ripples- like Rayleigh deep fadings [6] withsome periodicity across all subcarriers when thetransceiver devices were moving. Fig. 1 (a) and (b)show the CSI ripples in different 802.11n Modula-tion and Coding Scheme (MCS)∗ configurations.A natural question raised in our mind is “Are these

802.11n uses MCS value to represent different transmission variables combinations. Higher MCS value often denoteshigher rate. you may refer www.mcsindex.com for details.

Jizhong Zhao et al.: Indoor Localization with CSI-speed estimation 3

ripples correlated with distance or speed?”.

1 5 10 15 20 25 30

1

200

400

600

800

1000

Subcarriers

Sam

ples

0

5

10

15

20

25

30

1 5 10 15 20 25 30

1

200

400

600

800

1000

Subcarriers

Sam

ples

5

10

15

20

25

30

35

40

(a) Ripples in SISO (b) Ripples in MIMO

Fig. 1. (a) Ripples across nearly all subcarriers in 1× 1

SISO with MCS=0. (b) The ripples in 3 × 3 MIMOconfiguration with MCS=16.

The answer is beyond an intuitive yes. In thiswork, we first exhibit a simple yet precise relation-ship between moving speed v and the frequency ofripples-like fadings fr.

Based on this important discovery, we pro-pose a simple method to precisely estimate a wire-less device’s moving speed (or moving distance)purely based on wireless traffic. Since in Wi-Fienvironment both AP and clients can measure theCSI and RSS, the client’s moving speed and RSScan be measured remotely at AP-end passively andnon-intrusively. In such case, all connected devicesbecome the participants of the crowdsourcing sys-tem without any effort.

Based on the precise speed and moving dis-tance estimation, our localization scheme C2IL ap-plies the graph matching (GM) for several coretasks: Wi-Fi fingerprints extraction, automaticmapping between fingerprints and floor plan, andusers localization/tracking. It has been widelyknown that the RSS value is affected by many fac-tors, e.g., the RSS values collected at the samelocation using same devices with same WiFi APscould fluctuate to a few db depending on how usershold and block the signal [1]. Such fluctuationwill significantly impact the fingerprint matchingquality and thus impact the localization/trackingaccuracy. We first identify and successfully ad-

dress the directional shadowing problem for con-ventional threshold-based fingerprints extraction,and we propose a trajectory matching-based solu-tion to eliminate the shadowing problem. For fin-gerprints mapping task, C2IL supports unsuper-vised floor plan mapping for large scale complex in-door environment. For localization/tracking task,a combination of trajectories matching and particlefilter is proposed to achieve precise indoor local-ization/tracking. In summary, the contributionsof C2IL are as follows.

• Compared with the previous approacheswith good accuracy, our scheme C2IL does notrequire the localization devices to be smartphoneswith various inertial sensors (which are requiredin [2, 5]). As it only requires 802.11n connection,C2IL has the lowest barrier on client, thus, weexpect it to have better contribution from users,which is of vital importance for practical andcontinuously-functioning crowdsourcing system.

• C2IL is a practical localization system thatreally benefits from the multipath effect, insteadof suffering from this notoriously challenging ef-fect. Theoretical analysis showed that it is exactlythe multipath effect that enables the accurate dis-tance estimation by CSI. Our extensive experimen-tal evaluation indicated that in typical indoor sce-narios, the distance estimation error is often within3% regardless of moving speeds, which is muchmore accurate than Dead-Reckoning or pedometerbased approaches.

• The adoption of graph matching and othertechniques in our core design guarantees the accu-racy and scalability of RSS radio map and localiza-tion in very large and complex environment. Wedesign and develop a prototype of C2IL in a largeoffice environment of about 2000m2 with complexstructure. In our tests, the localization error with-out any historical data is within 5m; while in track-ing mode, the tracking error could be within 1m.


The rest of the paper is organized as follows.We review related techniques in Section 2, presentC2IL system overview in Section 3 and our innova-tive distance estimation in Section 4. Fingerprintextraction and the complete crowdsourcing basedlocalization scheme are introduced in Section 5 andSection 6 respectively. We present the fine-grainedindoor tracking solution in Section 7, and reportour extensive performance evaluation of C2IL inSection 8. We conclude the paper in Section 9.

2 Related Work2.1 Indoor Localization Schemes

The great success of GPS [7] doesn’t lead toa finish line for the localization problem. In bat-tery constraint wireless ad-hoc network, great ef-forts [8, 9] have been made to bring better accu-racy, and in indoor environment, the signal shad-owing and multipath effect also block the GPS orsimilar ideas.

fall into two categories: Fingerprint-based andModeling-based. Fingerprints are utilized in manyliteratures to assist positioning, and the mostwidely used one is WiFi signal. In indoor environ-ment, fingerprint based methods (e.g., Radar [10],Horus [11]„ SurroundSense [12], PinLoc [13], ) firstcollect fingerprint of WiFi signal (or cellular, orFM, or other sensors such as light) in advance atknown locations inside a building, and then iden-tify the user’s location by matching the fingerprintof this user with the fingerprint stored in database.Dead-Reckoning is another stream of techniques(e.g., [2, 14, 15]) proposed in the literature for lo-calization.

The mostly used fingerprint is the RSS value.LiFS [5] proposed a crowdsourcing based indoorlocalization, which exploits the possibility of au-tomatically establishing the mapping between fin-gerprint set F and position set P In our previ-ous work [16], graph matching algorithm is used

to achieve the automatic mapping in complex en-vironment. Acoustic ranging (AR) assisted Wi-Fipositioning was recently developed to provide dis-tance estimation between two users (e.g., [3, 4]).These schemes leverage the accurate AR and areable to provide high localization accuracy usingthe mapping of fingerprints with some additionaldistance constraint.

CSI has potential for accurate indoor local-ization since the CSI tool [17] has been released topublic on off-the-shelf hardware. CSI is not a sim-ple extension of RSS on physical subcarriers but itreveals totally different information on frequencyselective fading process [18]. SpinLoc [1] proposeda rotation based indoor localization system thatleveraged the human bodies’ strong fading to Line-of-Sight (LoS) components. PinLoc [13] proposeda CSI fingerprint-based localization system whichcan achieve meter-level precise indoor point lo-calization. Based on the strong position distinc-tion property of CSI, CSITE [19] proposed an Wi-Fi Management Frame Authentication framework.Besides the above device-based localization, Frog-Eye [20] proposed an CSI-based device-free crowddensity estimation system. Twins [21] proposedan critical state based device-free tracking system.Compared with these schemes, our scheme providean accurate moving speed estimation of a singleuser in a complex indoor environment.

Recently, the popularity of Software DefinedRadio (SDR) system brings more powerful ap-proaches. ArrayTrack [22] use an antenna ar-ray driven by SDR to provide accurate Angle-Of-Arrival (AoA) based indoor localization. On theother hand, Wi-See [23] and PinIt [24] simulatethe antenna array by a moving antenna. Thelatest work WiTrack [25] implements Frequency-modulated continuous-wave (FMCW) radar tech-nology on SDR platform, and it achieves 3Ddevice-free human tracking. Although SDR plat-


form exhibits strong potential, the deploymentcost and computation requirement are still toohigh for wide range deployment.

2.2 Estimate Moving Speed by WirelessSignal

There is a long history of estimating mov-ing velocity of a mobile station according to wire-less signal [26, 27]. Most of them focused on fastmoving stations, i.e., a mobile station in cars ortrains. However, the algorithm for estimating themaximum Doppler frequency fd, on which mostof these methods based, is not suitable for es-timating human walking speed. The maximumDoppler frequency in 2.4G or 5G Wi-Fi environ-ment could be almost totally ignored, i.e., fora moving station with a velocity 1.5m/s, fd ismerely 12Hz/25Hz compared to the carrier fre-quency 2.4Ghz or 5.2Ghz.

To the best of our knowledge, [28] is the onlyprevious work that implemented an indoor speedestimation system, which is based on DVB-T sig-nal working at 746Mhz. This work used the rela-tionship v = ξ λ

Tcto estimate velocity, where ξ =

0.423 a pre-defined constant, λ the wavelength,and Tc is the channel coherent time. However,for Wi-Fi signal with small covering range, whichcauses non-uniformly Rayleigh fading, a constantξ is not appropriated, and it is very difficult andchallenging to precisely estimate the ξ in a dy-namic indoor environment.

2.3 Graph Matching via Relaxation

Graph matching (GM) is a widely used tech-nique to find the best correspondence between twographs, documents, or images. In C2IL, we useGM to extract RSS fingerprints and automaticallyestablish the mapping between RSS fingerprintsand floor plan.

GM is essentially an integer quadratic pro-gramming (IQP) problem,and it is NP-hard [29].Given two graphs GP = (VP , εP ) and GQ =

(VQ, εQ), the goal of GM is to find the best cor-respondence between two graphs. Let CP×Q rep-resent the possible matching candidates set, theaffinities between all candidates are recorded inan adjacency matrix MCP×Q based on applica-tions. Let X ∈ {0, 1}nMCP×Q be a column-wise bi-nary vector which indicates the selected correspon-dences, the graph matching problem can be ex-pressed as finding the best indicator vector X ∗ thatmaximizes a score function S(X ) = X TMCP×QX ,i.e., X ∗ = argmaxS(X ).

Various relaxation based methods have beenproposed. Most of them relax the integer con-straint of X . After obtaining the optimal X ∗ inreal number domain with different insights of M,the discretization of X ∗ will make the best approx-imated solution to the underlined IQP problem. InC2IL, two major graph matching algorithms, spec-tral matching (SM) [30] and Reweighted RandomWalks Matching (RRWM) algorithms [29] are usedin different stages. SM is used in extracting RSSfingerprints graph; while RRWM is used in map-ping between RSS fingerprints and floor plan.

3 Architecture Overview

Since CSI and RSS can be estimated by bothAPs and clients, C2IL can be deployed in eitherAP-end or client-end. Whenever it is deployed ateither end, the core of the system remains same. Inthe rest of the paper, for simplicity of presentation,we assume that WiFi APs will collect the signaland conduct the needed computation. The posi-tions of WiFi APs are not required to be known ineither case.


��

��

��

��

��

!��

"�#��$�� $��

Fig. 2. C2IL System Architecture.

Fig. 2 demonstrates our system architecture.When a user enters the building, we assume thatthe device held by the user will build wireless con-nections with the APs inside the building. TheIEEE 802.11n WiFi APs will record the RSS val-ues and the CSI values of the received signal fromthe client device and send the data to a localizationserver. Based on the sequence of the CSI valuestaken during a time-window, the server will thenquickly estimate the moving distance of the clientin this time-window. The RSS values will be usedto build a matching with the fingerprint RSS val-ues stored in the fingerprint database, which itselfis populated using crowdsourcing techniques. Theestimated moving distance, together with the es-timated geodesic distance of different fingerprintlocations in the map, will be used to further im-prove the quality of the matching, and thus, theaccuracy of the localization. In the rest we discussin detail each of the components in C2IL.

4 Estimating Moving Speed by CSI

In this section, we mainly focus on the tech-niques of estimating moving speed and thus mov-ing distance by CSI. Theoretical basis is presentedfirst, and then we present our algorithm implemen-tation.

4.1 The Electromagnetic Standing WaveField

Wireless radio propagation in compact envi-ronment could be modelled as a superposition of

large-scale path-loss, middle-scale shadowing, andsmall-scale multipath fading [6]. For the multipathfading, it is usually fitted to a statistical modelcalled Rayleigh or Rician (Rayleigh fading plusstrong LoS components) distribution. The ripples-like deep fading shown in Fig.1 (b) and (c) aretypical Rayleigh fading pattern.

Previous speed estimation methods, are basedon some statistical properties of Rayleigh distribu-tion, e.g. level crossing rate (LCR) or coherenttime Tc. Although it has been experimentally val-idated that the distance between two adjacent rip-ples (deep fadings) is about λ/2 (λ is the carrierwavelength) even in large-scale multipath environ-ment (like Manhattan city) [31], no previous worksexplicitly exploit such λ/2 fluctuation, since suchfluctuation is encapsulated and blurred in a toogeneral model.

However, some detailed studies of radio prop-agation [32, 33] have indicated that, in a com-plex multipath environment the constructive or de-structive interferences of the large sum of reflectedand scattered waves will generate a standing wavesfield, and the environment becomes a weak Elec-tromagnetic Cavity Resonator (ECR) [34] whichhold standing waves in a very short time. Ac-cording to basic physics of wave propagation, thedistance between two adjacent antinodes (positionwith maximum amplitude), towards any direction,is λ/2, thereby the experimentally observed λ/2

fluctuation. Therefore, when an antenna traversesthe indoor space with a speed v, a periodicallyripples-like pattern with a frequency fo = 2 v

λ ap-pears. Such simple relationship inspires us thatthe moving speed v could be precisely estimatedpurely from the CSI, if we could precisely estimatethe fo.


4.2 Speed Estimation Algorithm

As previously described, the speed estimationproblem now becomes a specific frequency estima-tion problem. We design a reasonable and effec-tive processing flow. It includes Data Preprocess-ing, Noise Cancellation, Fading Enhancement, andFrequency Estimation.

Data Preprocessing: Every frame sent in802.11n MCS rate at time t has an CSI Ht. It is acomplex-number vector with a length of Nss×Lss,where Nss and Lss are the number of MIMO spa-tial streams and number of measured subcarriersacross the Wi-Fi bandwidth. Every complex valuehit ∈ Ht describes the instantaneous amplitudeait and phase θit of the underlying i-th subcarrier.In order to enable all 802.11n compatible devicesto be ready for speed estimation, we only use thefirst spatial stream (first Lss values in Ht) to es-timate speed. Moreover, the computation spaceis greatly reduced. Since in multipath environ-ment the phase θi is uniformly distributed between[0, 2π] [6], which provides no discriminative infor-mation. Thus we drop the phase θit and only useamplitude A = |H| to estimate speed.

The amplitude matrix Aori = {A1, ...,An}T

is further defined, where Ai is the ith receivedcolumn-wise amplitude. Since the instantaneousreception rate of frames is unstable due to the wire-less traffic control, Aori is resampled to a stablereception frequency fw with the even interval be-tween each slot, and let Are denote the resampledamplitude matrix.

Noise Cancellation: Convolution based noisecancellation is applied on Are to filter out the highfrequency noise, that Anc = Are ∗ hb(r),

hb(r) = r · 1s

where 1s is a full-1 square matrix of size s. Cur-rently in our system s = 6. This step is of greatimportance according to the real data evaluation,

since the following fading enhancement and fre-quency estimation is quite sensitive to noise.

Fading Enhancement: An intuitive idea of en-hance the fading is first-order derivation of Anc,however, first-order derivation is quite sensitive tohigh frequency noise rather than low frequency rip-ples. Another convolution is used to emphasize thefading that Aen = Anc ∗hdf , where hdf is a Sobel-style calculator that

hdf =

2 5 2

0 0 0

−2 −5 −2

(1)

Fig. 3 shows the intermediate results After first3 processing, and it is now suitable for frequencyestimation.

Am

plitu

de

Samples

50 100 150 200 250 300−20

0

20

40

60 Original Filtered Enhanced

Fig. 3. The effect of each step of processing.

Frequency Estimation: Due to the MIMO con-figuration or other interference, deep fading(s) inall subcarriers are not guaranteed to appear simul-taneously, as shown in Fig. 1, therefore the finaldecision of fo are based on the estimations of eachunderlying subcarriers.

Extracting f io for i-th subcarrier is equiva-

lent to extracting the expected frequency E(f ic)

in the spectral graph within a frequency intervalfmin < fc < fmax. According to Eq. (5), fmin

and fmax are set according to the speed interval ofhuman walking that

fmin = 2 · vmin/λ, fmax = 2 · vmax/λ (2)


where the minimum speed vmin and the maximumspeed vmax in our system are set to 0.8m/s and1.6m/s.

Short-Time Fourier Transformation (STFT)with 50% overlapping window is applied to obtainthe Power Spectral Density (PSD) of i-th envelopeof Are. It reveals the spectral density of subcar-rier i along with time. To reduce the jitter, theestimated f i

o is set to the weighed expectation offrequencies between fmin and fmax

f io =

∑fmin<sj<fmax

sj · wj∑wj

, (3)

where wj denotes the power of frequency fj .

Frequency (Hz)

Tim

e (s

)

0 25 50 75 1000

10

20

30

40

0

100

200

300

400

500

Fig. 4. STFT result for 15th subcarrier at 5.8Ghz(Channel 161).

Figure 4 shows the STFT result for 15-th sub-carrier. We can see very strong power around50Hz. As 50× λ

2 = 1.29m/s, and it is quite close tothe real walking speed at about 1.3m/s. Two redbars denote the fmin and fmax, and they are setto 32 and 64 according the vmin and vmax settings.The final estimation of fo is set to the median ofall estimated f i

o

fo = median(f1o , f

2o , ...f

no ), n = Nss (4)

Then the moving speed is estimated by

v =λ · fo2

(5)

It was worth to mention that although theDoppler effect does exist, as discussed in Section2, the Doppler frequency is small comparing to the

carrier frequency. Thus in current system design,we did not consider the Doppler effect caused byhuman walking.

4.3 Stop Detection

Precise Start and Stop detection of humanwalking by wireless device is a challenging prob-lem. Since instantaneous state shift of movingor static mainly reflect on small-scale fading, westill use CSI for stop detection. Fig. 5 (a) presentsan example when it starts moving at about 400thsample. Observing the degree of disorder beforeand after the start, an intuitive method is tocheck for the deviation of amplitude, however, thesliding-window required by deviation checking re-duces the checking sensitivity to movement. Fig. 5(b) shows the slow raising deviation values withdifferent length Lw of sliding window, and the de-tected time obviously lags behind.

The spatial de-correlation property of CSIgives us a hint. We find that the correlation coeffi-cient ρ between consecutive CSI samples will droprapidly if the spatial distance ds between them islarger than λ/2. Thus, there will be a rapid co-efficiency raising or drop to check the “moving”and “static” status. Fig. 5 (c) shows the sam-ples’ correlation matrix. When device is static,stable and high correlation co-efficiency holds theentire upper-left area, while it disappears immedi-ately when the device starts moving. Figure 5 (d)presents the moving averages of ρ with differentsliding window length Lw, and the dropping of ρ

is fast and clear. According to our experimentalevaluation, a devices is said to be moving when ρ

drops below 0.4, and the final detected time td isquite close to the actual time ta.


❙✉❜❝❛rr✐❡rs ❢r♦♠ ✸ ❛♥t❡♥♥❛s ✭✸ ★ ✸✵ ❂ ✾✵✮

Sam

ples

1 10 20 30 10 20 30 10 20 30

1

100

200

300

400

500

600

700

800 0

10

20

30

40

1 100 200 300 400 500 600 700 8000

1

2

3

4

5

Samples

Mov

ing

Avg

. of s

tand

ard

devi

atio

n

L

w=30

Lw

=70

Lw

=150

Actual Time

Detected Time

Samples

Sam

ples

1 100 200 300 400 500 600 700 800

1

100

200

300

400

500

600

700

800 −0.5

0

0.5

1

1 100 200 300 400 500 600 700 8000

0.2

0.4

0.6

0.8

1

Samples

▼♦✈✐♥❣❛✈❣✳♦❢❀

L

w=30

Lw

=70

Lw

=150

Detected TimeActual Time

(a) (b) (c) (d)

Fig. 5. (a) the CSI image of a sample. The movement starts at about 400 time-slot. (b) the moving averages ofstandard deviation with different window lengths Lw. The detected start delays about 100 slot than actual start.(c) the correlation matrix of the CSI images shown in (a). (d) the moving averages of the correlation coefficientswith different Lw.

4.4 Minimum Sampling Rate

Similar to sensor-based system, sufficient CSIsampling rate is critical for accuracy. Due tothe un-equal time distribution between fading andnon-fading, the Nyquist sampling rate of fs =

2 · fmax is not sufficient. We carried out experi-ments to find the minimum fs that can guaranteegood accuracy. Evaluations are carried out in awide range of channel frequencies including 2.4G(channel 1), 5.2Ghz (channel 40), 5.5G (channel100) and the highest 5.8G(Channel 161). Duringthe experiment, testers are walking at the samespeed around v = 1.3m/s and the mobile devicein their hands are constantly transmitting beaconframes at 500hz. After the experiments, we sim-ulate the sampling rate fs from 20hz to 500hz bydropping frames uniformly. Fig. 6 presents theresults. We can see from the figure that the es-timated speed v continuously climbs when fs ishigher than Nyquist rate fN , and the speed stopsraising when fs is about 4 times of fv = 2v/λ.More experiments in other situations have alsoconfirmed the 4 · fv sampling rate. Therefore if weset fmax = 1.6m/s, the minimum sampling rate isonly 100 (or 250) frames/s in 2.4G (or 5.8G) envi-ronment, or equivalent to approximately 40KBpsor 100KBps traffic.

20 60 100 140 180 220 260 300 500

0.9

1.0

1.1

1.2

1.3

Sampling Rate (Hz)

Mea

sure

d S

peed

(m

/s)

↑f2.4N

↓ f2.44x

←f5.2/5.5/5.8N

↑f5.24x

↓f5.54x

↑f5.84x

2.4GHz(Channel 1)5.2GHz(Channel 40)5.5GHz(Channel 100)5.8GHz(Channel 161)

Fig. 6. The measured speeds in different sampling rate.The points denoted by f⋆

N denotes the Nyquist sam-pling rate, while f⋆

4x denotes the minimum require sam-pling rate.

The traffic burstiness is another problem.The burstiness, which happened frequently, is ob-viously against the CSI-based speed estimation.Since the burstiness is usually short-time high-frequency traffic phenomenon, a reasonable as-sumption could be made to ease this problem: peo-ple’s walking speed remain stable during the gapbetween two burstiness. Fig. 7 presents our solu-tion that during each burstiness the speed is esti-mated, while in the gap the speed is approximatedas the average.


0 1 2 3 4 5 6 7 8 9 100

100

200

300

400

500

Time (s)

Fra

mes

Rat

e (f

/s)

1.201.221.241.261.28

Est

imat

ed S

peed

(m

/s)

Fig. 7. The burstiness of wireless traffic in practicalenvironment and the estimated speed.

5 RSS Fingerprints Extraction

RSS fingerprints are the most representativeRSS points for given positions, and the error ofthe fingerprint will directly affect the accuracy.Besides the strong noise, we observed the direc-tional shadowing problem which may easily breakthe traditional fingerprint extraction strategy.

5.1 The Directional Shadowing Problem

−86 −76 −66 −56 −46−82

−72

−62

−52

−46

RSS from AP1 (dBm)

RS

S fr

om A

P2

(dB

m)

−80

−70

−60

−50Same physical path

Fig. 8. RSS trajectories with 3APs. The color denotesthe value for 3rd AP: deeper color denotes higher RSSfrom 3rd AP.

When people walk along the same path in dif-ferent directions, the shadowing effect of humanbody will generate deviated RSS sample trajecto-ries. Fig. 8 presents an example of such direc-tional shadowing problem. Due to the strong di-rectional shadowing, the same physical path cor-responds to two parallel RSS trajectories. If ap-plying clustering-based fingerprint extraction, thealgorithm may wrongfully determine that there are

two parallel paths between the start and end point.Thus, eliminating such directional shadowing iscrucial for improving the localization accuracy.

5.2 Extracting Fingerprints

The RSS samples are highly deviated for thesame physical position if under directional shadow-ing, however, the whole RSS changing trends areremarkably similar. Therefore extracting uniqueRSS fingerprints without directional shadowingcould be transformed to a problem of correctlyidentifying the correspondence between a lot ofRSS trajectories. Treating these trajectories ascurve-shaped graphs, identifying the correspon-dence is, however, a n-partite graph matchingproblem. Fortunately, this can be done by n − 1

times iteratively graph matching between n-th tra-jectory and previous resulting RSS fingerprintsgraph. As revisited in Section 2, relaxation-basedapproaches approximate the graph matching ele-gantly, and the only job is to build the affinitymatrix MCP×Q for two graphs GP and GQ.

A possible graph matching between GP andGQ is a set of assignments( or pairs) (i, i′), wherei ∈ V P and i′ ∈ V Q. And for every two assign-ments a = (i, i′) and b = (j, j′), there is an affin-ity score ma,b that measures the compatibility be-tween them. Let set CP×Q enumerates all possibleassignments, then MP×Q is the affinity matrix ofit, and MP×Q stores all the affinity scores for allpossible assignments. Based on this definition, anoptimal graph matching is a set of assignments C∗

which has the highest inner affinity scores. Eachelement ma,b ∈ M is assigned as follows in current

For visual clarity, we plot only RSS traces from 3 APs.


system.

M(a,b) =

e−||dij−di′j′ ||, if ||dij − di′j′ || > ϵ

0, otherwise

(6)where dij and di′j′ are the distance between i andj, and their assignment pairs i′ and j′, respectively.Since RSS attenuation along distance is non-linear,in our system the Minkowski distance [35] withvalue p = 1.7 is used to define the distance be-tween RSS samples a and b.

dab = (n∑

i=1

|rssai − rssbi|p)1p (7)

where n represents the numbers of all heard APs,and rssai for the i-th AP’s RSS value of RSS sam-ple a.

1 100 200 300 4001

100

200

300

400

RSS sampling ID

RS

S s

ampl

ing

ID

0

0.5

1.0

Fig. 9. The correspondence ratio among RSS samplingsequence. Ted dots denote the extracted RSS finger-prints.

After solving the graph matching problemwith affinity matrix M using spectral approaches,the resulting column-wise vector X with lengthnP × nQ can then be reshaped to an associationmatrix AnP×nQ , where each element Aij denotesthe matching rate between i ∈ V P and j ∈ V Q.Due to the uniqueness, RSS fingerprints can beeasily extracted by finding the RSS points whichdo not have correspondence in historical data. LetUA be the upper triangular matrix of A and DA

for its diagonal matrix, a RSS point i is considered

to be a RSS fingerprint iff∑i

(UA −DA) = 0.

Fig. 9 shows a sample association matrix for thegraph matching between two same RSS trajecto-ries. The red dots denote the extracted uniqueRSS samples points by graph matching algorithms,and they become effective fingerprints.

5.3 Fingerprints Transition Graph

The Fingerprints Transition Graph GT

records the spatial connections of all fingerprints.There are two conditions that qualify two nodesi, j ∈ V (GT ) to have an edge eij ∈ ET :

1. i and j are subsequent RSS fingerprints vec-tor within the same RSS fingerprints seg-ment;

2. i and j belong to different segments, the dis-tance dij is smaller than ϵ and at least oneof both is the start or end of a segment.

The weight of each edge eij is set to the geodesicdistance between these two fingerprints measuredby CSI speed estimation module.

6 Mapping Between Fingerprints andFloor Plan

6.1 Floor Plan in Manifold’s Eyes

In human terms, the shortest distance be-tween two points i and j in the indoor environ-ment is not the euclidean distance deu(i, j) butthe geodesic distance dgeo(i, j), which means thatthe indoor space in locally euclidean but globallynot. In another words, indoor space could be es-sentially viewed as a 2D-manifold S embedded ina 2D-polygon P.

Since the geodesic distance reveals the truestructure of the manifold, we resample the 2Dfloor plan with uniformly scattered points and con-struct an n-nodes graph GM to represent the un-


folded version of the manifold. Any two nodesi, j ∈ V M have an edge eij iff the correspondencepoints i′, j′ in the floor plan P are in their mutualneighbourhood with direct line-of-sight distance,and the weight for edge is the direct distance thateij = deu(i

′, j′).

6.2 Unsupervised Accurate Mapping

Since the RSS samples are measured alongusers’ walking trajectories, the RSS fingerprintstransition graph GT also share the same mani-fold structure. An intuitive solution to establishthe fingerprint map is to apply graph matchingdirectly between GF and GM . However, the accu-racy and performance of large-scale graph match-ing (>50 points) is too poor for an unsupervisedsolution. A lightweight problem relaxation is toapply graphing matching only on corridor pointsgraph. Once the corridor points graph are mappedcorrectly, the room matching is trivial. Unfor-tunately, the accuracy and performance of graphmatching between such simple graphs is still notsatisfied.

We devise a method called “Skeleton Match-ing” to achieve unsupervised accurate mapping be-tween GF and GM even for very complex indoorstructure. We first performs clustering on the cor-ridor points to form a skeleton graph. Due tothe high sparsity chain structure of GF and GM ,the topology of the skeleton graphs follow exactlythe structure of corridor points. As a highly con-densed structure, performing graph matching onthese graphs could achieve very high accuracy.Finally, once the skeleton graphs are correctlymatched, the corridor points and the rooms willbe matched trivially. The algorithm is detailed infollowing 3 steps, including skeleton graphs extrac-tion, skeleton graphs matching, and find-grainedpoints matching.

Extract the skeleton graph

Two sub-steps are required to extract theskeletons and the skeleton graphs. 1). Identi-fying the corridor points graph GCF ∈ GF andGCM ∈ GM ; 2). extracting skeleton graphs GSF

and GSM .In the first sub-step, a customized centrality

measure C(V ) is designed to identify the core cor-ridor network. For a given point v ∈ V , its cen-trality C(v) is as follows.

C(v) =∑

s̸=v ̸=t∈Vσst(v) (8)

where σst(v) is the numbers of shortest path froms to t via v.

Based on this definition, we design a itera-tive algorithm to remove the non-central pointseffectively. In each round of iteration, the cen-trality C(v) is measured for every points. If C(v)

is smaller than a low-bound τ , then remove thepoints from the graph. This procedure repeats un-til no points is removed. The choose of value τ

is critical. Too small τ will bring in redundantpoints, while overlarge τ will make all points beremoved. Since the raise of τ will lead to mono-tonic decrease of number of remaining points, weuse binary search method to find the critical τ∗.

In the second sub-step, the skeleton graphsV S is generated by clustering the corridor pointsgraph GCF and GCM . We use spectral cluster-ing [36] (SC) as the clustering algorithm. SC iscomputationally faster than K-means and it onlyrequires the similarity matrix which is exactly suit-able in our case that both GCF and GCM are rep-resented only in adjacency matrix. By clusteringon GCF and GCM , we obtain the vertices set ofskeleton graph GSF and GSM . The edge set ESF

and ESM follow the underlying points, that if twopoints i, j, belonging to different clustersca and cb

respectively, have an edge, then there is an edge


between ca and cb. The weight of edge eab is de-fined as the shortest distance between the centralpoints of cluster a and b, and the central point ofa cluster is the point i which has the shortest dis-tances to other points within the cluster.

Skeletons Matching

After the extraction of the skeleton graphs forboth GF and GM , we then find the best correspon-dence between GSR and GSM via graph matching.Let symmetrical square matrix MSR and MSM

represent their adjacency matrices. We build theaffinity matrix MSR×SM for graph matching asfollows.

MSR×SM = e(1SM

⊗MSR−1SR

⊗MSM )

2

(9)

where⊗

denotes the Kronecker product [37] and1SR denotes the full-1 matrix with the same sizeof GSR.

We adopt RRWM algorithm [29] to performthe graph matching. Hungarian algorithm is fur-ther applied to discretize the X in order to meetthe final integer constraints X ∈ {0, 1}n.

Corridor Points Matching

Graph matching between GSF and GSM isquite easy. However, due to the inconsistency ofclustering on GCF GCM , the matching betweenthe skeletons graph cannot lead to the direct oneto one points matching of two matched clusters.

In the corridor points graphs, only a fewpoints connect multiple chain structures. There-fore, once we identify the correct correspondencesof these bridge points V CF

b and V CMb , the rest of

the points can be matched subsequently. In orderto identify the bridge node, we introduce a newmetric called ”bridge centrality”, which is equal tothe number of shortest paths from all vertices toall others within nearby clusters that pass through

that node.

Cbg(v) =∑

{(s,t)|v∈ci,s,t∈NN(ci),s,t/∈ci}

σst(v) (10)

where NN(ci) denotes the nearby cluster of ci.The bridge point will be the point with the highestbridge centrality.

After identifying the correspondence of bridgepoints in GCF and GCM , we next start the match-ing of the chain points in both graphes. Sincethe bridge points hold the main structure ofthe graphs, the enumeration of all the one-hoppaths between any two bridge points will cover allthe chains. For every one-hop path matched inboth graphs, the points in GCF are sequentiallymatched to GCF according to distance informa-tion.

7 Localization and Tracking

C2IL provides a unified localization andtracking service by treating the direction local-ization request as a tracking request without his-torical data. Here we mainly focus on track-ing technique in C2IL. Unlike the stateless K-NNbased method which is widely adopted in previ-ous approaches, in our solution, the users’ trajec-tories are globally determined from the very be-ginning by transforming the tracking problem toa graph matching problem between the measuredRSS samples transition graph GS and the finger-print transition graph GT . After the graph match-ing, the accuracy is further improved by bring-ing in the CSI-based speed estimation through aparticle-filter based fusion. Here we start intro-ducing these two steps.

7.1 Graph Matching Based Tracking

Graph matching based tracking is to find thebest correspondence between the sequence of RSS


samples of tracking request and extracted finger-prints. This is exactly the same matching processundertaken in Section 5, except for the differencesthat the tracking is to find the matched points,while in Section 5 the un-matched RSS samplesare added to the fingerprints database.

Let X nS×nF represent the association matrixobtained through spectral matching where nS andnF are the numbers of RSS samples and candi-date fingerprints respectively. Due to the errorin RSS measurement and fingerprints map con-struction, a single RSS samples pi may corre-spond to multiple fingerprints in X ,e.g., a RSSsample may correspond to two fingerprints, oneis in corridor, and another is in a room. Fortu-nately the temporal correlation can help eliminatethose false correspondence by checking the spatialcontinuity between current and subsequent candi-date fingerprints. After eliminating the false corre-spondence, the globally estimated coordinates se-quence will be given by TGS = fGT (GS), wheref : fingerprint → GM represents the mappingfrom fingerprints to the floor plan manifold GM ,and TGS are the resulting coordinates sequence inGM .

However, full graph matching between all fin-gerprints set and RSS samples is time-consuming,therefore a matching candidates pruning process isintroduced to meet practical demand of real-timetracking of multiple clients.

Matching Candidates Pruning: The mainidea of the pruning is to find the probable walkingarea for the tracking request using coarse-grainednearest neighbor (NN) method, which will signifi-cantly reduce the search space.

Let RT = {rT1 , .., rTn } represent the RSS se-quence of the tracking request at time t, where ri

is a n-dimensional RSS vector. The length n issmaller than a positive integer L, such that thetracking algorithm provides limited backtracking.

Let RF = {rF1 , .., rFn } denote the RSS fingerprintsset and graph GF

P denote the positions of finger-prints in the floor plan. The pruned matching can-didate set Su is defined as follows.

Su = {rF |rF ∈ RF , rF ∈ ϵ-NNRF(NNRF(RT))}(11)

where NNRF(RT) denotes the nearest neighbors ofRT within the fingerprints set RF , and ϵ-NNA(B)

denotes the B’s neighborhood within distance ϵ inthe set A. Here the ϵ is set to 3m.

7.2 Fine-Grained User Tracking

The movement estimation from trajectoriesmatching (TM) and CSI-based speed estimation(CBSE) are naturally complementary. TM is slow-responsive yet accurate in long time, while CBSEmay drift with time cumulation yet be accurate inshort time. Obviously, a fusion scheme for TMand CBSE will significantly improve the accuracy.Due to the linearity constraints and difficulties incorrectly estimating the error covariance matrices,we don’t use the conventional Kalman filter. Themore robust particle filter (PF) is adopted as thefusion algorithm.

The state space of the tracking is a two di-mensional vector Xt = [Mi, vt], where Mi rep-resents the i-th node in the floor plan manifoldM , and v for the walking speed. The main chal-lenge here is that the speed v has no direction in-formation, therefore every round the PF evolvesand there are multiple candidate predictions. Forinstance, when walking in a corridor without op-tional paths, there are two candidate directions,forward and backward, and therefore two statecandidates Xfor = [Mj , vt] or Xrev = [Mh, vt],where Mj and Mh are two nearby nodes in differ-ent directions.

Fortunately,the PF can handle this problemelegantly by sprinkling different amounts of parti-


cles in different directions. In the prediction phaseof each filter iteration, the particles will be re-sampled and enumerate all possible candidate di-rections.

Fig. 10 illustrates the particles distributionalong a path, in each round the particles enumer-ate all possible candidate directions. When thereare optional paths, the particles will enumerate allpossible options. After the fusion, the Mi is trans-formed to floor plan coordinates.

Fig. 10. The predicted particles distribution before eachnew measurement: red dots represent the particles, andgreen dots for the determined position in last round.

8 System Evaluation

8.1 Prototype System with CustomizedAPs

We designed and developed a prototype hard-ware system for C2IL. The prototype system isdeployed in a large 2000m2 office environmentwith circular corridor network as shown in Fig. 11.To maximize the participation rate, C2IL is de-ployed at AP-end. In total, 19 customized APsare sparsely deployed across the office that pro-vide both wireless networking service and localiza-tion service. They will forward the measured RSSand CSI values to a central localization server. Adeveloped simple client App is installed on testers’Android smartphones. The localization server willcalculate the clients’ position based on the net-working traffic, and clients only need to read thelocation coordinate from server.

0 10 20 30 40 50 60 67.80

5

10

15

20

25

29

Length (m)

Wid

th (

m)

Fig. 11. The floor plan of our test area, and red dotsdenote the deployed customized APs.

Fig. 12. Prototype system for C2IL with Intel Atom-based Mini PC and 5300NIC as AP.

Although CSI is a standard PHY layer infor-mation, currently only Intel 5300 NIC can exportit to user level. Our customized AP is simply an In-tel ATOM- based mini PC with 5300 NIC. Figure12 shows our customized AP. It equips with single-core 1.6Ghz ATOM CPU and 5300 NIC. The totalcost is about $90. The OS is Ubuntu 12.04 64-bit,and AP function is hosted by hostapd. Besides theordinary AP interface, a monitor virtual interfaceis also added to overhear the wireless traffic. Bothmeasured CSI and RSS are uploaded to localiza-tion server in real-time.

8.2 C2IL Speed and Distance Estimation

In this experiment, 10 students are asked towalk 3 times around the 129m long circular corri-dor as shown in Fig. 11. They are asked to main-tain constant walking speed in first round, and thespeed may change slightly and remarkably in sec-ond and third round. The measured speed will


be integrated to walking distance Dw. We mainlyconsidered the error rate ew = |Dpath−Dw|/Dpath.

Two other speed estimation approachesi.e. pedometer-based and channel coherent-timebased [28], are also developed as comparison. Forpedometer we use NASC [2] method to detectsteps, and the step length is predefined in train-ing phase. The accelerometer data required by thealgorithm was retrieved from laptop’s HD activeprotection. For coherent-time based approaches,we use a constant ξ = 0.396, which is manuallyoptimized for this evaluation.

1% 3% 5% 7% 9% 11% 13% 15%0

0.2

0.4

0.6

0.8

1

CD

F o

f Spe

ed E

stim

atio

n E

rror

Error Rate

R1.CSIR2.CSIR3.CSIR1.ChntR2.ChntR3.ChntR1.PedoR2.PedoR3.Pedo

Fig. 13. The Error CDF of 3 methods.

Fig. 13 presents the error CDF of three meth-ods in three walking manners: with constant speed(R1), slightly varying speed (R2), and with markedchange (R3). The experiment shows that, com-paring to pedometer- based approach, C2IL withCSI can achieve better accuracy without requir-ing predefined constants or training data, mean-while, C2IL significantly outperforms pedometer-based (Pedo) approaches when speed is varyingor pre-defined step length is out of effectiveness.Coherent-time based approach (Chnt) haves thesimilar speed-invariance feature, however, the ac-curacy is considerably poorer than C2IL and pe-dometer in constant speed.

The accuracy of CSI-based distance estima-tion is then evaluated. To evaluate the influence ofrichness of multipath components to the distanceestimation error, we carried out experiments in 3typical environments, a compact corridor, a large

office environment, and a very large hall. In eachenvironment we walked along a 50m straight linefor 10 times. Two APs were simultaneously usedto estimate walking distance. One was placed atthe end of the path with strong LoS component,and the other was placed in a cubicle to cut-offthe LoS components to simulate strong Rayleighfading. Figure 14 (a) plots the Rician K Fac-tor [38] along the walking path in different envi-ronment, which estimate the degree of LoS com-ponents. Very low and stable K appears in corri-dor environment which means there are rich multi-path components, while in office and hall the mul-tipath components is significantly reduced due tothe weak reflection in large wide-open space. Fig-ure 14 (b) and (c) plot the CDF of estimated walk-ing distance by the AP in path and cubicle respec-tively. We see in the best situation that in a cor-ridor with strong multipath component, there isonly 3% error. In Figure 14 (c), we also see smallerror happened in the path, and in the worst sit-uation, in a large hall with very weak multipathcomponents, the averaged estimated error is lessthan 10%.

8.3 Mapping Accuracy Test

Fig. 15 (a) shows extracted corridor pointsand the ”bridge points”. The critical τ∗ is set to127 after 6 round of iteration. To obtain the RSStransition graph, 5 students are asked to walkingin the rooms for 20 minutes to cover all acces-sible area. The RSS and CSI data is collected,and Fig. 15 (b) and (b) shows the visualizationof GT and GCT respectively. Please note thatour mapping algorithm requires only the adjacencymatrix of graphs, and the MDS-based position isonly for visualization. The graph matching resultis shown in figure (d). For each graph GSF orGSM there are 8 points, the graph matching undersuch small scale could achieve stably and accurate


0 10 20 30 40 500

100

200

300

400

500

Walking Distance (m)

❊st✐♠❛t❡❞❘✐❝✐❛♥❋❛❝t♦r❑

HallOfficeCorridor

45 46 47 48 49 50 51 520

0.2

0.4

0.6

0.8

1

Estimated Distance (m)

CD

F R

ate

HallOfficeCorridor

45 46 47 48 49 50 510

0.2

0.4

0.6

0.8

1

Estimated Distance (m)

CD

F R

ate

HallOfficeCorridor

(a) Racian Factor K (b)Strong Rayleigh Fading (c) Weak Rayleigh Fading

Fig. 14. (a) the Rician factor K along the path in different environment. (b) the CDF of estimated distancewithout LoS components (c) the CDF with strong LoS components.

0 10 20 30 40 50 60 67.80

5

10

15

20

25

29

Length (m)

Wid

th (

m)

(a) (b) (c)

Floor Plan ClustersFingerprints Clusters

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

Error (m)

Err

or C

DF

Corridor PointsRoom Points

(d) (e) (f)

Fig. 15. (a) The clustering of corridor points graph GCM , the purple points denotes the ”bridge points”. (b) Anvisualization of RSS fingerprints transition graph GT . (c) The clustering of the GCT (the corridor points of GT ).(d) The skeleton matching result between GST and GSM . (e) The matching between GCT and GCM under theguide of skeleton matching and the ”bridge points”, the same color denotes the correspondence. (f) The roommatching accuracy CDF.

mapping. Figure (e) shows the matching of cor-ridor points under the guide of skeleton matchingand the ”bridge points”. Figure (f) shows the errorof both corridor points mapping and room pointsmapping. We can see clear that the corridor pointsare highly matched. The corridor points maximumerror is still under 2m, while the maxmium roommatching error is under 6m.

8.4 Localization and Tracking

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

Err

or C

DF

Localization Error (m)

RADARRADAR Tracking (6m)

C2IL(3m)

SENIL(3m)

C2IL(6m)

SENIL(6m)

Fig. 16. The localization & tracking error CDF with dif-ferent amount of historical RSS fingerprints data. TheRADAR scheme is as comparison. The suffix num-ber (0m) denotes the direct single point localization,while (10m) denotes the tracking service with a histor-ical data of 10m.


In this test, we compared C2IL with our pre-vious work SENIL [16], and RADAR [10] systemstill worked as baseline. the RSS/CSI samples of20 minutes walking in different speeds are used astesting data. Since we mainly evaluated the ac-curacy with different length of historical data, werandomly select ℓ successive RSS samples for eachround. Figure 16 plots the error CDF with differ-ent amount of historical data. Comparing to SE-NIL, C2IL is averagely 1m better. The accuracyimprovement is mainly due to the precise movingspeed/distance estimation.

9 Conclusion

In this paper, we proposed an indoor local-ization and tracking scheme, C2IL. Our schemedoes not require using additional sensors, exceptthe availability of 802.11n wireless connection. Webelieve that C2IL is the first scheme that the reallybenefits from the multipath effect in complex en-vironment. An innovative method is proposed toaccurately estimate the moving speed and distancepurely based on 802.11n CSI, which should find awide range of applications alone. Based on this ac-curate distance estimation, we built the mappingbetween RSS fingerprints and location using unsu-pervised learning, and unified the localization andtracking. Our extensive evaluation results indicatethat our scheme C2IL successfully handles verycomplex indoor structure and simultaneously pro-vides the best performance in contribution rate,localization cost, and localization/tracking accu-racy.

References

[1] Sen S, Choudhury R R, Nelakuditi S. Spinloc:Spin once to know your location. Proceedingsof the Twelfth Workshop on Mobile Comput-ing Systems & Applications, 2012, p.12.

[2] Rai A, Chintalapudi K K, Padmanabhan V N,Sen R. Zee: zero-effort crowdsourcing for in-door localization. Proceedings of the 18th an-nual international conference on Mobile com-puting and networking, 2012, pp.293–304.

[3] Nandakumar R, Chintalapudi K K, Padman-abhan V N. Centaur: locating devices in anoffice environment. Proceedings of the 18thannual international conference on Mobilecomputing and networking, 2012, pp.281–292.

[4] Liu H, Gan Y, Yang J, Sidhom S, Wang Y,Chen Y, Ye F. Push the limit of wifi based lo-calization for smartphones. Proceedings of the18th annual international conference on Mo-bile computing and networking, 2012, pp.305–316.

[5] Yang Z, Wu C, Liu Y. Locating in finger-print space: wireless indoor localization withlittle human intervention. Proceedings of the18th annual international conference on Mo-bile computing and networking, 2012, pp.269–280.

[6] Rappaport T S, et al. Wireless communica-tions: principles and practice. Prentice HallPTR New Jersey, 1996.

[7] Liu J, Priyantha B, Hart T, Ramos H S,Loureiro A A, Wang Q. Energy efficient gpssensing with cloud offloading. Proceedings ofthe 10th ACM Conference on Embedded Net-work Sensor Systems, 2012, pp.85–98.

[8] Zhao J, Xi W, He Y, Liu Y, Li X Y, Mo L,Yang Z. Localization of wireless sensor net-works in the wild: pursuit of ranging qual-ity. IEEE/ACM Transactions on Networking(TON), 2013, 21(1):311–323.

[9] Bo C, Ren D, Tang S, Li X Y, Mao X, HuangQ, Mo L, Jiang Z, Sun Y, Liu Y. Locating sen-


sors in the forest: A case study in greenorbs.INFOCOM, 2012 Proceedings IEEE, 2012,pp.1026–1034.

[10] Bahl P, Padmanabhan V N. Radar: An in-building rf-based user location and trackingsystem. INFOCOM, 2012 Proceedings IEEE,2000, pp.775–784.

[11] Youssef M, Agrawala A. The horus loca-tion determination system. Wireless Net-works, 2008, 14(3):357–374.

[12] Azizyan M, Constandache I, Roy ChoudhuryR. Surroundsense: mobile phone localiza-tion via ambience fingerprinting. Proceedingsof the 15th annual international conferenceon Mobile computing and networking, 2009,pp.261–272.

[13] Sen S, Radunovic B, Choudhury R R, MinkaT. You are facing the mona lisa: spot localiza-tion using phy layer information. Proceedingsof the 10th international conference on Mo-bile systems, applications, and services, 2012,pp.183–196.

[14] Constandache I, Choudhury R R, Rhee I. To-wards mobile phone localization without war-driving. INFOCOM, 2010 Proceedings IEEE,2010, pp.1–9.

[15] Guha S, Plarre K, Lissner D, Mitra S, KrishnaB, Dutta P, Kumar S. Autowitness: locatingand tracking stolen property while toleratinggps and radio outages. ACM Transactions onSensor Networks (TOSN), 2012, 8(4):31.

[16] Jiang Z, Zhao J, Han J, Tang S, Zhao J, XiW. Wi-fi fingerprint based indoor localizationwithout indoor space measurement. MobileAd-Hoc and Sensor Systems (MASS), 2013IEEE 10th International Conference on, 2013,pp.384–392.

[17] Halperin D, Hu W, Sheth A, WetherallD. Tool release: gathering 802.11 n traceswith channel state information. ACM SIG-COMM Computer Communication Review,2011, 41(1):53–53.

[18] Yang Z, Zhou Z, Liu Y. From rssi to csi: In-door localization via channel response. ACMComputing Surveys (CSUR), 2013, 46(2):25.

[19] Jiang Z, Zhao J, Li X Y, Han J, Xi W.Rejecting the attack: Source authenticationfor wi-fi management frames using csi infor-mation. INFOCOM, 2013 Proceedings IEEE,2013, pp.2544–2552.

[20] Xi W, Zhao J, Li X Y, Zhao K, Tang S, Liu X,Jiang Z. Electronic frog eye: Counting crowdusing wifi. 2014.

[21] Han J, Qian C, Ma D, Wang X, Zhao J, ZhangP, Xi W, Jiang Z. Twins: Device-free objecttracking using passive tags. 2014.

[22] Xiong J, Jamieson K. Arraytrack: a fine-grained indoor location system. Usenix NSDI,2013.

[23] Adib F, Katabi D. See through walls withwifi! Proceedings of the ACM SIGCOMM2013 conference on SIGCOMM, 2013, pp.75–86.

[24] Wang J, Katabi D. Dude, where’s my card?:Rfid positioning that works with multipathand non-line of sight. Proceedings of the ACMSIGCOMM 2013 conference on SIGCOMM,2013, pp.51–62.

[25] Adib F, Kabelac Z, Katabi D, Miller R C. 3dtracking via body radio reflections. 2014.

[26] Azemi G, Senadji B, Boashash B. Mobile unitvelocity estimation based on the instanta-


neous frequency of the received signal. Vehic-ular Technology, IEEE Transactions on, 2004,53(3):716–724.

[27] Mohanty S. Vepsd: a novel velocity estima-tion algorithm for next-generation wirelesssystems. Wireless Communications, IEEETransactions on, 2005, 4(6):2655–2660.

[28] Pricope B, Haas H. Experimental valida-tion of a new pedestrian speed estima-tor for ofdm systems in indoor environ-ments. Global Telecommunications Confer-ence (GLOBECOM 2011), 2011 IEEE, 2011,pp.1–5.

[29] Cho M, Lee J, Lee K M. Reweighted ran-dom walks for graph matching. In ComputerVision–ECCV 2010, Springer, 2010, pp.492–505.

[30] Leordeanu M, Hebert M. A spectral techniquefor correspondence problems using pairwiseconstraints. Computer Vision, 2005. ICCV2005. Tenth IEEE International Conferenceon, 2005, pp.1482–1489.

[31] Sklar B. Rayleigh fading channels in mo-bile digital communication systems. i. charac-terization. Communications Magazine, IEEE,1997, 35(7):90–100.

[32] Sheriff I, Belding-Royer E. Multipath selec-tion in multi-radio mesh networks. Broad-band Communications, Networks and Sys-tems, 2006. BROADNETS 2006. 3rd Interna-tional Conference on, 2006, pp.1–11.

[33] Yang S C. OFDMA system analysis and de-sign. Artech House, 2010.

[34] Hill D A. Electromagnetic fields in cavities:deterministic and statistical theories. JohnWiley & Sons, 2009.

[35] Groenen P J, Jajuga K. Fuzzy clustering withsquared minkowski distances. Fuzzy Sets andSystems, 2001, 120(2):227–237.

[36] Von Luxburg U. A tutorial on spectralclustering. Statistics and computing, 2007,17(4):395–416.

[37] Weichsel P M. The kronecker product ofgraphs. Proceedings of the American Mathe-matical Society, 1962, 13(1):47–52.

[38] Doukas A, Kalivas G. Rician k factor es-timation for wireless communication sys-tems. Wireless and Mobile Communications,2006. ICWMC’06. International Conferenceon, 2006, pp.69–69.

Dr. Jizhong Zhao He isa Professor at the Departmentof Computer Science and Tech-nology, Xi’an Jiaotong Univer-sity. His research interests in-clude computer software, per-vasive computing, distributedsystems, network security. Heis a member of CCF, ACM, andIEEE.

Zhiping Jiang is a Ph.Dcandidate at Xi’an JiaotongUniversity, Xi’an. His researchinterests include localization,smart sensing, wireless commu-nication, and image process-ing. He is a student memberof IEEE.

Dr. Xiang-Yang Li is aprofessor at the Illinois Insti-tute of Technology. He holdsEMC -Endowed Visiting Chair


Professorship at Tsinghua Uni-versity. He currently is distin-guished visiting professor at Xi’An JiaoTong University, Uni-versity of Science and Technol-ogy of China, and TongJi Uni-

versity. His research interests include mobile com-puting, cyber physical systems, wireless networks,security and privacy, and algorithms. He is a se-nior member of IEEE and a member of ACM.

Dr. Shaojie Tang is anassistant professor at the Uni-versity of Texas at Dallas. Hereceived his Ph.D degree fromIllinois Institute of Technology.His research interests includesmart grid, sensor network, al-gorithm, and wireless security.

Dr. Jinsong Han iscurrently an associate professorat Xi’an Jiaotong University.He received his Ph.D. degreefrom Hong Kong University ofScience and Technology. Hisresearch interests include per-

vasive computing, distributedsystem, and wireless network.He is a member of IEEE and

ACM.

Dr. Wei Xi is a postdoc-toral research fellow at Xi’anJiaotong University. His mainresearch interests include wire-less networks, smart sensing,and mobile computing. He isa member of CCF, ACM, andIEEE.

Kun Zhao is an Ph.Dcandidate at Xi’an JiaotongUniversity, Xi’an. His researchinterests include signal process-ing, wireless communication,smart sensing, and algorithm.

Zhi Wang is an Ph.D can-didate at Xi’an Jiaotong Uni-versity, Xi’an. His research in-terests include smart sensing,pervasive computing, localiza-tion, and wireless security.

Communicating Is Crowdsourcing: Wi-Fi Indoor Localization ...

Documents

Transcript of Communicating Is Crowdsourcing: Wi-Fi Indoor Localization ...