IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 1

CSI-based Indoor LocalizationKaishun Wu Member, IEEE Jiang Xiao, Student Member, IEEE, Youwen Yi Student Member, IEEE,

Dihu Chen , Xiaonan Luo and Lionel M. Ni Fellow, IEEE

Abstract—Indoor positioning systems have received increasing attention for supporting location-based services in indoor environ-ments. WiFi-based indoor localization has been attractive due to its open access and low cost properties. However, the distanceestimation based on received signal strength indicator (RSSI) is easily affected by the temporal and spatial variance due to themultipath effect, which contributes to most of the estimation errors in current systems. In this work, we analyze this effect across thephysical layer and account for the undesirable RSSI readings being reported. We explore the frequency diversity of the subcarriers inOFDM systems and propose a novel approach called FILA, which leverages the channel state information (CSI) to build a propagationmodel and a fingerprinting system at the receiver. We implement the FILA system on commercial 802.11 NICs, and then evaluate itsperformance in different typical indoor scenarios. The experimental results show that the accuracy and latency of distance calculationcan be significantly enhanced by using CSI. Moreover, FILA can significantly improve the localization accuracy compared with thecorresponding RSSI approach.

Index Terms—Indoor localization, Channel State Information, RSSI, Physical Layer.

F

1 INTRODUCTION

LOcalization is one of the essential modules of manymobile wireless applications. Although Global Posi-

tioning System (GPS) works extremely well for an open-air localization, it does not perform effectively in indoorenvironments due to the disability of GPS signals topenetrate in-building materials. Therefore, precise in-door localization is still a critical missing componentand has been gaining growing interest from a widerange of applications, e.g., location detection of assetsin a warehouse, patient tracking inside the building ofthe hospital, and emergency personnel positioning in adisaster area.

A great number of researches have been done toaddress the indoor localization problem. Many range-based localization protocols compute positions based onreceived signal strength indicator (RSSI), which repre-sents the received power level at the receiver. Accordingto propagation loss model [1], received signal powermonotonically decreases with increasing distance fromthe source, which is the foundation of the model-basedlocalization. Most of the existing radio frequency (RF)-based indoor localization are based on the RSSI val-

• K. Wu is with National Engineering Research Center of Digital Life, State-Province Joint Laboratory of Digital Home Interactive Applications, SunYat-sen University, Guangzhou, China. He is also affiliated at Fok YingTung Graduate School, HKUST. Email:{kwinson}@ust.hk.

• Jiang Xiao, Youwen Yi and Lionel M. Ni are with Department of Comput-er Science and Engineering, HKUST. E-mail: {jxiao,ywyi,ni}@cse.ust.hk.

• Dihu Chen is with School of Physics and Engineering, Sun Yat-senUniversity, Guangzhou, China. E-mail: {stscdh}@mail.sysu.edu.cn.

• Xiaonan Luo is with National Engineering Research Center of DigitalLife, State-Province Joint Laboratory of Digital Home InteractiveApplications, Sun Yat-sen University, Guangzhou, China. E-mail:{lnslxn}@mail.sysu.edu.cn.

ues [1]–[5]. More related work is in the supplementalfile. However, we claim that the fundamental reasonswhy RSSI is not suitable for indoor localization arefrom two aspects: First, RSSI is measured from the RFsignal at a per packet level, which is difficult to obtainan accurate value. According to our measurement ina typical indoor environment as shown in Fig. 1, thevariance of RSSIs collected from an immobile receiver inone minute is up to 5dB. Second, RSSI is easily varied bythe multipath effect. In theory, it is possible to establisha model to estimate the separation distance using thereceived power. In reality, however, the propagation ofa RF wave is attenuated by reflection when it hits thesurface of an obstacle. In addition to the line-of-sight(LOS) signal, there are possibly multiple signals arrivingat the receiver through different paths. This multipatheffect is even more severe in indoor environments wherea ceiling, floor and walls are present. As a result, it ispossible for a closer receiver to have a lower RSSI thana more distant one. Consequently, a simple relationshipbetween received power and separating distance cannotbe established. Therefore, this time-varying and vulner-able RSSI value creates undesirable localization errors.

We argue that a reliable metric provided by commer-cial NICs to improve the accuracy of indoor localiza-tion is in need. Such metric should be more temporalstable and provide the capability to benefit from themultipath effect. In current widely used Orthogonal Fre-quency Division Multiplexing (OFDM) systems, wheredata are modulated on multiple subcarriers in differentfrequencies and transmitted simultaneously, we have avalue that estimates the channel in each subcarrier calledChannel State Information (CSI). Different from RSSI,CSI is a fine-grained value from the PHY layer whichdescribes the amplitude and phase on each subcarrier inthe frequency domain. In contrast to having only one

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 2

RSSI per packet, we can obtain multiple CSIs at onetime. More importantly, the CSIs over multi-subcarrierswill travel along different fading or scattering pathson account of the multipath effects. It then naturallybrings in the frequency diversity attribute of CSI, whicheach subcarrier has different amplitudes and phases. Byexploiting the frequency diversity, we can construct aunique “fingerprinting” indicating each location on theradio map. According to these advantages, it is favorableto leverage the CSI to improve the performance of indoorlocation fingerprinting. And thus, designing a precisetracking/localization system becomes possible.

Based on CSI, in this paper, we present the design andimplementation of FILA, a novel cross-layer approachbased on OFDM for indoor localization using WLANs.

In summary, the main contributions of this paper areas follows.

1) We design FILA, a cross layer approach that en-ables fine-grained indoor localization in WLANs.FILA includes two parts, the first one is CSI-basedpropagation model and the second one is CSI-based fingerprinting. To the best of our knowledge,FILA is the first to use fine-gained PHY layerinformation (CSI) in OFDM to build a propagationmodel so as to improve indoor localization perfor-mance. And it is also the first time to take advanceof the combination of the fine-grained PHY lay-er information CSI with frequency diversity andmultiple antennas with spatial diversity for indoorlocation fingerprinting.

2) We implement FILA in commercial 802.11 NICsand conduct extensive experiments in several typ-ical indoor environments to show the feasibility ofour design.

3) Experimental results demonstrate that FILA sig-nificantly improves the localization accuracy ascompared to the corresponding traditional RSSI-based approach.

The rest of this paper is organized as follows. InSection. 2, we introduce some preliminaries. This isfollowed by the system architecture design in Section 3.In Section 4, we demonstrate the CSI-based propagationmodel. In Section 5, we illustrate the methodology ofCSI-based fingerprinting. The implementation of FILAand experimental evaluations are presented in Section 6.Finally, conclusions are presented and suggestions aremade for future research in Section 7.

2 PRELIMINARIESIn this section, we introduce the CSI value which isthe foundation of FILA design. And the backgroundinformation of the OFDM system can be found in sup-plemental file.

2.1 Channel State InformationBased on OFDM, channel measurement at the subcarrierlevel becomes available. Nowadays, adaptive transmis-sion systems in wireless communication always improve

−34 −32 −30 −28 −26 −240

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

RSS(dBm)

Pro

bab

ility

Fig. 1: Temporal variance of RSSI.

the throughput by utilizing some knowledge of the chan-nel state to adapt or allocate transmitter resources [6].

Channel state information or channel status informa-tion (CSI) is information that estimates the channel byrepresenting the channel properties of a communica-tion link. More specifically, CSI describes how a signalpropagates from the transmitter(s) to the receiver(s) andreveals the combined effect of, for instance, scattering,fading, and power decay with distance. In summary,the accuracy of CSI greatly influences the overall OFDMsystem performance. It is worth pointing out that accord-ing to the definition of CSI, only OFDM-based WLANsystems can demonstrate the frequency diversity in CSIsince they use multiple subcarriers for data transmission.In another word, other modulation schemes like DSSScannot provide this value.

In a narrowband flat-fading channel, the OFDM sys-tem in the frequency domain is modeled as

y = Hx+ n, (1)

where y and x are the received and transmitted vectors,respectively, and H and n are the channel matrix andthe additive white Gaussian noise (AWGN) vector, re-spectively.

Thus, CSI of all subcarriers can be estimated accordingto (1) as

H =y

x, (2)

which is a fine-grained value from the PHY layer thatdescribes the channel gain from TX baseband to RXbaseband.

3 SYSTEM DESIGN

In this section, we first give an overview of the systemarchitecture. Challenges in the system design are pre-sented in the supplemental file.

3.1 System ArchitectureFILA system is built based on the current communi-cation system and thus compatible to the under layerdesign. More precisely, no modification is required at

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 3

+

CSI

OFDM

Decoder

CSIeff

Localization Block

Normal

Data

(2) Process CSI

Mobile

TX RXDistance

RF Signal PropagationStationary

RX

AP Location

InformationTX

(2)’ Distance

Calculator(3) Locate RX

(1) Collect CSI

OFDM

Demodulator

X

X

Channel

Estimation

Fig. 2: System Architecture.

the transmitter end (TX–the AP), while only one newcomponent for CSI processing is introduced for localiza-tion purposes at the receiver end (RX–the target mobiledevice). Fig. 2 demonstrates the detailed design of thesystem architecture. For traditional packet transmissionin wireless communication, only the demodulated signalis exported to the decoder for message content retrieval.However, a prerequisite in FILA localization system isthat it should be able to export the CSI value after thenormal demodulation process. Such that we devise alocalization block to exploit the CSI information.

In our designated localization block, the CSI collect-ed from 30 groups different subcarriers will firstly beprocessed. After running the proposed algorithm, wecan obtain the effective CSI in an efficient time con-straint. Then the effective CSI will be used to estimatethe location of the target object. As mentioned in theprevious section, CSI value is the channel matrix fromRX baseband to TX baseband which is needed for chan-nel equalization. Therefore, there is no extra processingoverhead when obtaining the CSI information. Never-theless, RSSI is obtained at the receiver antenna in the2.4 GHz radio frequency before down convert to the IFand baseband. Therefore, the free space model that builtfor RSSI-based localization approaches can’t be directlyapplied to process the CSI value. We need to refine suchradio propagation model according to the CSI informa-tion and compute the distance based on the proposedone. Finally, as the AP location information is obtainedfrom the network layer while CSI is collected from thephysical layer, we then use the simplest trilaterationmethod to obtain the location. For the fingerprinting, weleverage the CSI values including different amplitudesand phases at multiple propagation paths, known as thefrequency diversity, to uniquely manifest a location. Wethen present a probability algorithm with a correlationfilter to map object to the fingerprints.

In our FILA system architecture design, CSI is onlyprocessed by a newly designed localization block if

needed. Owing to the fact, FILA can be applied con-currently with the original packet transmission. In otherwords, it will not introduce additional overhead duringthe data transmission.

4 CSI-BASED PROPAGATION MODEL

In this section, we leverage the fine grained CSI value in-stead of RSSI to build a propagation model and addressthe indoor localization issue. The CSI-based propagationmodel can be built based on three following steps.

1) CSI Processing: First, we need to mitigate estima-tion error by effectively processing the CSI valuedenoted as CSIeff . This is known as the prerequi-site of the ongoing two steps.

2) Calibration: Afterwards, we develop a refined in-door propagation model and a fast training algo-rithm to derive the relationship between CSIeffand distance.

3) Location Determination: By receiving the APs coor-dinates in network layer and CSIeff values fromphysical layer, we apply the revised propagationmodel and trilateration method to accomplish thelocalization. This part is in the supplemental file.

4.1 CSI ProcessingFor wireless communication, attenuation of signalstrength through a mobile radio channel is caused bythree nearly independent factors: path loss, multipathfading, and shadowing. The path loss characterizes theproperty that the signal strength decays as the distancebetween the transmitter and receiver increases, which isthe foundation of our CSI-based localization. Multipathfading is a rapid fluctuation of the complex envelope ofreceived signal caused by reception of multiple copiesof a transmitted signal through multipath propagation.Shadowing represents a slow variation in a receivedsignal strength due to the obstacles in propagation path.Therefore, before establishing the relationship betweenCSI and distance, we need to mitigate the estimationerror introduced by multipath fading and shadowing.

4.1.1 Time-domain Multipath MitigationThe first concentration of our design is that the sys-tem must be capable of dealing with the challenge ofoperating over a multipath propagation channel. Sincemultipath effect will introduce Inter-Symbol-Interference(ISI), cyclic prefix (CP) is added to each symbol tocombat the time delay in OFDM systems. However,the CP technique is helpless for the multiple reflectionswithin a symbol time.

For narrow-band systems, these reflections will not beresolvable by the receiver when the bandwidth is lessthan the coherence bandwidth of the channel. Fortu-nately, the bandwidth of 802.11n waveforms is 20MHz(with channel bonding, the bandwidth could be 40MHz),

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 4

which provides the capability of the receiver to resolvethe different reflections in the channel. We propose amultipath mitigation mechanism that can distinguishthe LOS signal or the most closed NLOS from otherreflections in the expectation of reducing the distanceestimation error.

The commonly used profile of multipath channel inthe time domain is described as follow,

h(τ) =

Lp−1∑k=0

αkδ(τ − τk), (3)

where Lp is the number of multipath channel componen-t. αk and τk are the amplitude and propagation delay ofthe k-th path.

In practice, OFDM technologies are efficiently imple-mented using a combination of fast Fourier Transform(FFT) and inverse fast Fourier Transform (IFFT) blocks.The 30 groups of CSI represent the channel response infrequency domain, which is about one group per twosubcarriers. With IFFT processing of the CSI, we canobtain the channel response in the time domain, i.e.,h(t). From Fig. 3 we can observe that the LOS signal andmultipath reflections come with different time delay, andgenerally the LOS signal has higher channel gain, so wecan use a trunk window with the first largest channelgain in the center to filter out those reflections. If LOSdoesn’t exist, we can identify the shortest path NLOS re-flection. According to Nyquist sampling theorem, widerspectrum leads to higher resolution in the time domain.Due to the bandwidth limitation of WLAN, we can’tdistinguish all the reflections but we can use this methodto reduce the variance induced by multipath effects. Thetime duration of the first cluster is determined by settingthe truncation threshold as 50% of the first peak value.In doing so, we expect to mitigate the estimation errorintroduced by multipath reflection.

After the time domain signal processing, we reobtainthe CSI using FFT. Fig. 3 shows the CSI results after timedomain filtering. Note that, commercial NICs embedshardware circuits for the FFT and IFFT processing, ouralgorithm introduces ignorable latency to the wholelocalization procedure.

4.1.2 Frequency-domain Fading CompensationMoreover, since CSI represents the channel responsesof multiple subcarriers, a combination scheme is alsointroduced to process the CSI value in our system forcompensation of the fading of received signals in fre-quency domain to enhance location accuracy.

In general, when the space between two subcarriersis larger than the coherence bandwidth, they are fadingindependently. Since the channel bandwidth of 802.11nsystem is larger than the coherence bandwidth in typicalindoor environment, the fading across all subcarriers arefrequency-selective. To combat such frequency selective-ly fading of wireless signals, multiple uncorrelated fad-ing subchannels (multiple frequency subcarriers), that is

−30 −20 −10 0 10 20 3010

20

30

40

50

60

Subcarrier Index

CS

I Am

plitu

de

Original valueAfter filtering

Fig. 3: Time Domain Channel Response.

30 groups of CSI values are combined at the receiver.Motivation for leveraging the frequency diversity stemsfrom the fact that the probability of simultaneous deepfading occurring on multiple uncorrelated fading en-velopes (in our case, resulting from frequency diversity)is much lower than the probability of a deep fadeoccurring on a single frequency system. Thus, exploitingthe wide bandwidth of WLAN that assures sufficientlyuncorrelated subcarriers, will reduce the variance in CSIsowing to small scale factors, which appears to be one ofthe major sources of location determination error. In ourFILA system, we weighted average the 30 groups CSIsin frequency domain so as to obtain the effective CSI,which exploits the frequency diversity to compensate thesmall-scale fading effect.

Given a packet with 30 groups of subcarriers, theeffective CSI of this packet is calculated as

CSIeff =1

K

K∑k=1

fkf0

× |Hk|, k ∈ (−15, 15), (4)

where f0 is the central frequency, fk is the frequency ofthe k-th subcarrier, and |Hk| is the amplitude of the k-thsubcarrier CSI.

Note that selection of weighting factors are basedon the fact that the radio propagation is frequency-related. According to the free space model, the receivedsignal strength is related to the frequency the signal istransmitted. So by this weighting method, we transferthe channel gain from multiple subcarriers to a singlesubcarrier, i.e., the central one. Next, we will establishthe relationship between the CSIeff and distance.

4.2 Calibration

Since CSI value is obtained from the baseband on thereceiver side, the radio propagation model [7] for RSSIis no longer suitable for our design. So we developa refined indoor propagation model to represent therelationship between the CSIeff and distance by revising

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 5

the free space path loss propagation model, given by:

d =1

4π

[(c

f0 × |CSIeff |

)2

× σ

] 1n

, (5)

where c is the wave velocity, σ is the environmentfactor and n is the path loss fading exponent. Both ofthe two parameters are dependent on distinctive indoorenvironments. The environment factor σ represents thegain of the baseband to the RF band at the transmitterside, inversely, the gain of RF band to baseband at thereceiver side, and the antenna gains as well. Moreover,for NLOS AP, the σ also includes the power loss due towall penetration or the shadowing. The path loss fadingexponent n is varying depending on the environment.For instance, when the RF signal is propagating along afree space like corridor, the path loss fading exponentn will be around 2. In other cases, such as an officethat represents a complex indoor scenario, the exponentcould be larger than 4. In an indoor radio channel withclutter in medium, where often the LOS path is aug-mented with the multipath NLOS at the receiver, signalpower decreases with a pass loss fading exponent higherthan 2 and is typically in the order of 2 to 4 [8]. Hence, itis not trivial to determine the received signal power andwe need to refine the free space propagation model thatobeys the analytical and empirical methods. A widelyused simplification is to assume that all the path lossexponents that model propagations between the specificreceiver and all the APs are equal. This simplificationin a typical indoor environment is an oversimplification,since the channel propagation is usually very differentdepending on the relative position of the mobile clientwith regard to each AP. Therefore, we calibrate bothenvironment factor σ and the path loss fading exponentn in a per-AP manner.

We propose a simple fast training algorithm basedon supervised learning to retrieve the parameters withthree anchors in offline phase. In the first step, CSIs ofmultiple packets are collected at two of the anchors totrain the environment factor σ and the path loss fadingexponent n for the refined indoor propagation model. Inthe second step, CSIs collected at the third anchor areused to test the efficiency of the parameter estimation.The two steps run iteratively until convergence. Theexperimental results in the next section show that thissimple algorithm can achieve satisfactory accuracy, moresophisticated training method will be able to obtainbetter performance.

5 CSI-BASED FINGERPRINTING

In this section, we introduce the methodology of CSI-based location fingerprinting approach. We start by us-ing a mobile device equipped with 802.11 NICs to re-ceive the beacon message from nearby APs at each sam-ple position. The message contains CSI that representsthe channel response of multiple subcarriers. We modify

the driver and divide the CSIs into 30 groups. Hence,N = 30 groups CSI values are collected simultaneouslyat the receiver that represented as

H = [H1,H2, · · · ,Hi, · · · ,HN ]T , i ∈ [1, 30], (6)

where each subcarrier Hi is defined as

Hi = |Hi|ej sin{∠Hi}, (7)

where |Hi| is the amplitude response and ∠H is thephrase response of the ith subcarrier.

Then, it comes to CSI processing which is the prerequi-site of calculating position likelihood in the positioningphase. We quantify the power of a package, denoted assummational CSI, by adding up the power with respectto 30 groups of subcarriers. Specifically,

He =

I∑i=1

|Hi|2, i ∈ [1, 30], (8)

5.1 CSI-based Fingerprinting GenerationAs the foundation of fingerprinting approach, the mea-sured CSI values are processed to construct a radiomap. Since most of the RF-based fingerprint methodsconsider two spatial dimensions for localization [9], wealso follow the principle. Therefore, the two-dimensionphysical space coordinate of a sample position lj islj = (lj,x, lj,y). To generate a radio map, we first extractthe statistic determine the number of detectable APsfor a sample position. At each reference point, we willestimate the signal strength distribution for each accesspoint at each location. In our location system, the signalstrength over all the subcarriers is represented by He.

Moreover, another component of the radio map willbe normalized CSI values at each subcarrier for eachAP. The motivation of leveraging the frequency diversitystems from the fact that CSI benefits from the mul-tipath effects, because the received signal at differentpositions will be the combination of different reflections.Therefore, these normalized CSI values will reflect thecombination result ignoring the large scale fading.

5.2 Positioning PhaseFor object location estimation, the target is required tobe accurately mapped to the radio map.

Previous works show that the probabilistic approachessuch as maximum likelihood provide more accurateresults than deterministic ones do in indoor environ-ments [9]. Therefore, we adapt the probability modelin [10] except that we use He instead of RSS value.Similarly, we treat He observed from the AP to thereceiver at a fixed location as a Gaussian variable. In theproposed system, we will select K best APs to calculatethe probability of the MS at each reference point. Thecriteria for the best AP selection is that those APs withhighest He values, because they are more reliable. In ourexperiment, we fix the K to be 3.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 6

The selected K He values obtained by the terminal tobe located form a vector He = [He,1, · · · ,He,K ]. Then,the position estimation problem is equivalent to findingthe l that maximizes the posteriori probability P (lj |He).According to Bayes’ law,

P (lj |He) =P (lj)P (He|lj)∑i P (li)P (He|li)

=P (lj)P (He|lj)

P (He)(9)

Note that P (lj) is the prior probability that the termi-nal located at the reference point li. In [9], [10], uniformdistribution is assumed. In contrast, we will leverage thespatial correlation of the CSI to determine the P (lj).

Recall that CSI is a fine-grained information, we canobserve channel response over multiple subcarriers rep-resented by Hk = [H1,H2, · · · ,Hi, · · · ,HN ]T for the kthAP. We denote the observed CSI with normalization foreach AP as H(O) ∈ CN×K , and the CSI recorded in theradio map for the same set of APs at position lj as H(lj).To quantify the similarity of the observed CSI and thestored “fingerprints” for all the APs, we use the Pearsoncorrelation between them which is defined as

ρH(O),H(lj) =K∏

k=1

cov(Hk(O),Hk(lj))

σHk(O)σHk(lj)

, (10)

where each AP is considered to be independent. Ac-cording to the measurement, the spatial channel cor-relation will decrease as the distance between the tworeceiver increases. Therefore, with higher ρ, the positionof the terminal will be closer to the reference point. Then,the probability of the terminal on each candidate pointis defined as

P (lj) =ρH(O),H(lj)∑Ji=1 ρH(O),H(li)

(11)

where J is the size of the candidate reference points set.Considering uncorrelated property between each AP,

the likelihood P (He|lj) can be calculated as,

P (He|lj) =K∏

k=1

P (He,k|lj), (12)

Since the signal strength at each reference point ismodeled as a Gaussian variable which requires less sam-plings than the histogram approach [11]. At the offlinephase, we can obtain the expectation He,k and varianceσe,k corresponding to the He,k, and the P (He,k|lj) isobtained as

P (He,k|lj) =1√

2πσe,k

exp−(He,k − He,k)

2

2σe,k. (13)

The location estimation of the terminal is the weightedaverage over the whole candidate set,

l =J∑j

P (lj |He)lj (14)

For the fingerprinting method, the terminal can pro-cess CSI by itself and then check the globe or local

database for localization. It doesn’t need to rely on onepubic server.

The performance of the proposed fingerprintingmethodology is evaluated in the following section.

6 EXPERIMENTAL RESULTS

In this section, we present the implementation and ex-perimental evaluation of FILA. First, we describe theexperimental setup which can be found in the supple-mental file. Then we illustrate the validation results forour refined propagation model. Finally, we evaluate theperformance of CSI-based propagation model and its fin-gerprinting. In our evaluation, we use the performanceof corresponding RSSI-based approach based on radiopropagation model and trilateration as baseline .

II. Experimental ScenariosWe conduct experiments to show the performance androbustness of our FILA system in four different scenariosin the campus of Hong Kong University of Science andTechnology as follows:

1) Chamber First, we set up a testbed in a 3m × 4mChamber to collect the RSSI and CSI as shown inFig. 4. In general, Chamber is an enclosure thatused as environmental conditions for conductingtesting of specimen. In our experiment, chamberrepresents the ideal free space indoor environmentwhich means only LOS signal exists without othermultipath reflection or external interference.

2) Research Laboratory Then, we deployed FILA inan identical indoor scenario – a 5m × 8m researchlaboratory as shown in Fig. 5. In the laboratoryregion, we place three APs on the top of three shel-ters in three dimensions. The experiment was con-ducted on a weekday afternoon when there werea couple of students sitting or walking around,which will show the robustness of our system totemporal dynamics of the environment. The laptopwas placed at a fixed position at the beginning ofthe experiment and then moved to the next pointalong a pre-measured path.

3) Lecture Theatre In addition, we chose a largerlecture hall to conduct the localization experiments,which is a 20m × 20m lecture theatre. Since thespace is relatively large, the influence of the roomsize can be explored.

4) Corridor Finally, we performed experiments in acorridor environment with multiple offices asidein our academic building, which is 32.5m × 10mcovering corridors, rooms and cubicles. In thisscenario, we expect to illustrate the impact of theabsence of LOS APs on the location accuracy.

6.1 Validate the Refined ModelAs the target for precise indoor localization, two mostimportant metrics are used to testify FILA: the accuracy

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 7

Fig. 4: Chamber Fig. 5: Research Laboratory

2.5 3 3.5 4 4.5 5 5.5 610

15

20

25

30

35

40

45

50

Distance (meters)

CS

Ieff

ampl

itude

CSIeff amplitudeExponential Fitting

Fig. 6: Relation between CSIeff andDistance.

0 10 20 30 40 50 601011121314151617181920

Time Duration(s)

CS

Ieff

ampl

itude

Lecture Theatre

Fig. 7: CSIeff on Temporal Variance

0 10 20 30 40 50 60−35−34−33−32−31−30−29−28−27−26−25

Time Duration(s)

RS

SI(

dBm

)

Lecture Theatre

Fig. 8: RSSI on Temporal Variance

0 2 4 6 8 10 12 14 160

0.5

1

1.5

2

2.5

3

Time Duration(min)

Err

or (

m)

CSIChamber

CSILab

RSSIChamber

RSSILab

Fig. 9: CSI vs. RSSI on distance esti-mation in different environments.

and the temporal stability of location estimation. After-wards, we compared the performance of our CSI-basedlocalization system with the corresponding RSSI-basedapproach.

6.1.1 Robustness of the Refined ModelOne essential aspect that needs to be determined beforethe localization experiments is whether CSI value canbuild a relationship with distance. In general, indoorRF signal strength is a non-monotonic function withdistance due to multipath and shadowing effects. Fig. 6illustrates the CSI value approximated by a power func-tion of distance according to our refined propagationmodel. In diverse scenarios with corresponding environ-ment factor σ, the path loss fading exponent n varies ina range of [2, 4]. It is shown that our refined model prop-erly fits the relationship between CSIeff and distance.

6.1.2 Temporal Stability of CSITemporal stability is a fundamental criteria in validatingthe robustness of the localization systems. We thus setout to examine the stability of the proposed new metricCSIeff and RSSI value in time series. It is well-knownthat RSSI is a fickle measurement of the channel gainbecause of its coarse packet-level estimation and easilyvaried by multipath effect. As CSI is fine-grained PHYlayer information that provides detailed channel stateinformation in subcarrier level, it is of great importanceto figure out whether it will remain in a stable mannerin practical environment.

Fig. 7 and Fig. 8 illustrate both the interactions ofCSIeff and RSSI values on temporal variance, respec-tively. It is shown that the received signal power cal-culated by RSSI has a variance up to 5 dBm, whilethe variance of CSIeff is within 1 dBm. Therefore,the proposed new matric is much more stable in timedomain compared with RSSI.

In Fig. 9, we further investigate the CSIeff valueand RSSI value in the chamber and research laboratoryso as to discover the effect of any temporal instabilityon distance estimation. Chamber provides a free space-like environment as it uses specific material that canabsorb the non-LOS signals. Thus, multipath effect canbe eliminated in chamber environment. However, theresult from our experiment shows that even in chamberthe RSSI is also varied significantly from time to timedue to the inaccurate measurement. In contrast, researchlab is a typical multipath environment. Both the staticobstacles and dynamic walking around individuals exertthe influence on multipath and bring in more intensepath loss. In this way, the variance of RSSI becomes evenlarger and the performance of distance estimation is evenworse.

Fig. 7, Fig. 8 and Fig. 9 lead to an essential conclusion:in comparison with RSSI, CSI is more temporally stablein different environments and helps maintain the per-formance over time. Therefore, FILA can achieve accu-rate location more quickly than the RSSI-based scheme,which is very crucial for some location-based application

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 8

RSSI CSI0

0.5

1

1.5

2

2.5

3

3.5

4

RSSI vs CSI

Mea

n di

stan

ce e

rror

(m

)

ChamberLaboratoryLecture Theatre

Fig. 10: Mean distance error.

like search and rescue.

6.2 Performance Evaluation of CSI-based Propaga-tion ModelAs a target for precise and fast indoor localization, twomost important metrics are used to testify this model: theaccuracy and the latency of localization, afterwards, wecompared the performance of our CSI-based localizationsystemwith the corresponding RSSI-based approach byusing the propagation model.

6.2.1 AccuracyAccuracy over a Single LinkAs the premise of indoor localization, we first inves-tigated the distance determination accuracy of FILAcompared with the corresponding RSSI-based approach.The primary source of error in indoor localization ismultipath propagation caused by multiple reflectionsthat overlap with the direct LOS subcarrier at the re-ceiver. FILA takes advantage of the fine-grained traitto mitigate such multipath effect, and exploits the fre-quency diversity to compensate the frequency-selectiveshading. We repeated the distance measurement experi-ments across 10 different locations in chamber, researchlaboratory and lecture theatre, respectively. For somepositions with serious multipath effect, FILA achieves upto 10 times accuracy gain over the corresponding RSSI-based scheme. Fig. 10 illustrates the mean distance errorsin three different environments. Our evaluation showsthat FILA can outperform the corresponding RSSI-basedscheme by around 3 times for the distance determinationof a single link.

To assess the effectiveness of the CSI-based localiza-tion approach, in the following we evaluate the accuracyof FILA in different typical indoor environments.

Localization Accuracy in Single RoomIn the experiments conducted in the research laboratory,we fix three APs on the top of the shelters. The mobilelaptop with iwl5300 NICs is first fixed at one locationand then moved to another. We repeated this process andplaced the device at 10 different positions respectively.Fig. 11(a) illustrates the cumulative distribution (CDF) of

localization errors across the 10 positions. In our experi-ments, for over 90% of data points, the localization errorfalls within the range of 1 meter, and the 50% accuracyis less than 0.5 m. In such a dynamic environment withlots of factors interfering the propagation of signals,FILA exhibits a preferable property indicating that thefine-grained nature of CSI is beneficial to improve theaccuracy of corresponding RF-based approach.

Fig. 11(b) depicts the cumulative distribution functionerrors per location detected at the university lecturetheatre. Even for such a much larger space, FILA canlocate objects in the range within 1.8 meters of theiractual position with 90 percent probability, which isacceptable for most location-based applications.

Across the above two typical single room indoor sce-narios, FILA achieves median accuracy of 0.45 m and 1.2m, respectively. It is therefore safe to conclude that theproposed CSI-based scheme performs much better thanRSSI-based one when locating objects in a typical indoorbuilding with multipath effect.

Localization Accuracy in Multiple RoomsIn our previous experiments, the APs and client areplaced in the same room. We also examined the corridorscenario where several APs are deployed in the multiplerooms. Specially, we take into consideration both thecomplicated multipath effect and the shadowing fadingbrought by wall shield. We first fix the position of theobject at some reference nodes, and collects the APcoordinate and CSI value for offline training. Then, wemove the object to arbitrary positions for online tracking.The moving speed is around 1m/s and we collect 20 CSIsand RSSIs at each position.

In Fig. 11(c), we plot the cumulative distribution oflocation errors across 10 positions. It is shown thatmultipath propagation does degrade the accuracy ofobject localization as well as the shadowing in the mul-tiple rooms scenarios. However, FILA is robust enoughto maintain the degradation. More importantly, FILAcan achieve median accuracy of 1.2 m in this corridorenvironment. This result indicates that FILA is able toeffectively estimate and compensate for gain differencesacross multiple rooms.

6.2.2 Latency of CSI-based propagation model localiza-tionTwo main phases contribute to the latency of FILA,the calibration phase and location determination phase.Since the environment factor and the fading exponentvary in different environments, we need to conductcalibration to train these two parameters for the refinedpropagation model. We should collect CSIs at some pre-known positions to calculate these two parameters usingour fast training algorithm. Actually, this process canbe finished before localization as an offline task sincethe APs can use each other’s information for the cali-bration. In our FILA system, the AP takes about 0.8msto transmit a packet with 100 bytes beacon message in

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 9

0 0.25 0.5 0.75 1 1.25 1.5 1.75 20

0.10.20.30.40.50.60.70.80.9

1

Distance error (m)

Pro

babi

lity

FILARSSI−based

(a) Research Laboratory

0 0.5 1 1.5 2 2.5 3 3.5 40

0.10.20.30.40.50.60.70.80.9

1

Distance error (m)

Pro

babi

lity

FILARSSI−based

(b) Lecture Theatre

0 0.5 1 1.5 2 2.5 3 3.5 40

0.10.20.30.40.50.60.70.80.9

1

Distance error (m)

Pro

babi

lity

FILARSSI−based

(c) Corridor

Fig. 11: CDF of localization error in different indoor environments.

IEEE 802.11n. Each time we collect 20 CSIs and the timewill be 0.8 × 20 = 16ms. The calibration process can bedone within 2ms according to our measurement on aHP laptop with 2.4GHz dual-core CPU. In the locationdetermination phase, the IFFT and FFT process can lever-age the according hardware blocks in the wireless NICswhose running time is ignorable. While we conductthese signal processing on laptop consuming around2ms, including the time needed for the trilaterationlocation calculation. Therefore, the time consumptionfor both training and location determination is withinseveral ms. In our experiment, the walking speed of useris around 1m/s. The average tracking error is around1.2m as shown in Figure 11(c). For multiple rooms envi-ronment, the simple alpha-tracking algorithm [12] can beapplied for triggering the training. We only need to trainthese parameters once unless the environment changesgreatly. In summary, our system can reach the averagetime tracking latency to as fast as about 0.01s, whichsignificantly outperforms previous RSSI-based trackingsystems [4] (usually 2− 3s).

1 2 3 4 5 60

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Number of APs

Mea

n di

stan

ce e

rror

(m

)

CSI−basedHorus

Fig. 12: Mean distance error.

6.3 Performance Evaluation of CSI-based Finger-printing6.3.1 AccuracyFirst, we evaluate the accuracy of the proposed CSI-based probability algorithm and compared with Horus,

the widely used RSSI-based fingerprinting system. Themean distance error in the two different environmentsis shown in Fig. 12. Our evaluation shows that CSI-based method can achieve the median accuracy of 0.65m,which outperforms Horus by about 0.2m, and the gainis about 24%. Moreover, in the corridor scenario, wherecovered by 6 APs and 3 APs were taken into computa-tion, the mean accuracy of our approach is 1.07m whichis 0.35m lower than Horus system, about 25% gain overHorus.

In addition, we compare the two approaches con-cerning different numbers of APs in corridor. Fig. 14depicts the average accuracy according to the amountof APs varying from 1 to 6. Since richer information toestimate the location can be obtained from the more APs,both lines demonstrate the accuracy improvement. Inparticular, our approach reduced the mean distance errorby 29% on the average. Obviously, these results show theeffectiveness of the proposed CSI-based location systemand indicate the benefits from indoor environment withdense-deployed APs. When the AP is sparse, our schemeperforms much better than the RSS-based one.

6.3.2 PrecisionFig. 15 illustrates the cumulative distribution (CDF) oflocalization errors in the laboratory. The data were col-lected across the 28 positions in the laboratory. FromFig. 15, we can observe that the confidence interval withconfidence level 90% for the error is (-1.3m, +1.3m),which means the CSI-based localization error falls withinthe range of 1.3 meters, and the 50% accuracy is less than0.6 m. However, the Horus can locate objects in the rangewithin 1.6 meters of their actual position with 90 percentprobability, and the median accuracy is 0.8.

Unlike the first scenario that 3 APs and client areplaced in the same room, we also examined the corridortestbed where the 6 APs are deployed in the multiplerooms. Fig. 13 depicts the cumulative distribution ofpositioning errors across 20 positions. We can easily ob-serve that both our approach and Horus can achieve themedian accuracy less than 1.25m. However, the accuracyimprovement of our approach over Horus for 90% of

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 10

0 0.5 1 1.5 2 2.5 30

0.10.20.30.40.50.60.70.80.9

1

Distance error (m)

Pro

babi

lity

CSI−basedHorus

Fig. 13: CDF of localization error inCorridor.

1 2 3 4 5 60

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Number of APs

Mea

n di

stan

ce e

rror

(m

)

CSI−basedHorus

Fig. 14: Mean distance error with dif-ferent numbers of APs.

0 0.25 0.5 0.75 1 1.25 1.5 1.75 20

0.10.20.30.40.50.60.70.80.9

1

Distance error (m)

Pro

babi

lity

CSI−basedHorus

Fig. 15: CDF of localization error inLaboratory.

data points is 0.55m. We can conclude that our approachexhibits a preferable property since the fine-grained andfrequency diversity nature of CSI is beneficial to improvethe precision of location fingerprinting.

7 CONCLUSIONS AND FUTURE WORKLocalization is one of the most appealing applicationsand becomes increasingly common in our daily life.RSSI-based schemes have been widely used to providelocation-aware services in WLAN. However, in this pa-per, we observe that RSSI is roughly measured and easilyaffected by the multipath effect which is unreliable. Wethen use the fine-grained information, that is, ChannelState Information (CSI), which explores the frequencydiversity characteristic in OFDM systems to build theindoor localization system FILA. In FILA, we process theCSI of multiple subcarriers in a single packet as effectiveCSI value CSIeff , and develop a refined indoor radiopropagation model to represent the relationship betweenCSIeff and distance. Based on the CSIeff , we thendesign a new fingerprinting method which leveragesthe frequency diversity. To demonstrate the effectivenessof FILA, we implemented it on the commercial 802.11nNICs. We then conducted extensive experiments in typ-ical indoor environments and the experimental resultsshow that the accuracy and speed of distance calculationcan be significantly enhanced by using CSI.

In this work, we just use the simplest trilaterationmethod to illustrate the effectiveness of CSI in indoorlocalization. The future research in the new and largelyopen areas of wireless technologies can be carried outalong the following directions. First, we can leverage theavailable multiple APs to improve the location accuracyin some extent. Second, in this paper we only leveragethe frequency diversity, however, the spatial diversitycan also be exploited in MIMO to enhance the indoorlocalization performance. Third, since some of the smartphones have 802.11n chipset, the next step of our workis to implement FILA in smart phone.

ACKNOWLEDGMENTSThis research was supported in part by the 2012Guangzhou Pearl River New Star Technology Train-ing Project, Hong Kong RGC Grants HKUST617811,

617212, S&T Project of Guangdong Province, ChinaGrant No.2011A011302001, NSFC Grant No. Grant No.(60933011 and 61027009), NSFC-Guangdong Joint Fund(U0835004), the National Key Technology R&D Program(2011BAH27B01, 2011BHA16B08), the Project Scienceand Technology of Guangdong Province (2011168014and 2011912004) and the Major Science and TechnologyProjects of Guangdong (2011A080401007).

REFERENCES[1] P. Bahl and V. N. Padmanabhan, “Radar: an in-building rf-based

user location and tracking system,” in Proc. of IEEE INFOCOM,2000.

[2] J. Liu, Y. Zhang, and F. Zhao, “Robust distributed node localiza-tion with error management,” in Proc. of ACM MobiHoc, 2006.

[3] D. Moore, J. Leonard, D. Rus, and S. Teller, “Robust distributednetwork localization with noisy range measurements,” in Proc. ofACM Sensys, 2004.

[4] D. Zhang, J. Ma, Q. B. Chen, and L. M. Ni, “An rf-based system fortracking transceiver-free objects,” in Proc. of IEEE PerCom, 2007.

[5] D. Zhang and L. M. Ni, “Dynamic clustering for tracking multipletransceiver-free objects,” in Proc. of IEEE PerCom, 2009.

[6] D. Halperin, W. J. Hu, A. Sheth, and D. Wetherall, “Predictable802.11 packet delivery from wireless channel measurements,” inProc. of ACM SIGCOMM, 2010.

[7] A. Goldsmith, Wireless Communications and Networks:3G and be-yond. Tata McGraw Hill Education Private Limited, 2010.

[8] K. Fazal and S. Kaiser, Multi-carrier and spread spectrum systems :from OFDM and MC-CDMA to LTE and WiMAX. Wiley, 2008.

[9] M. Youssef and A. Agrawala, “The horus wlan location determi-nation system,” in Proc. of ACM MobiSys, pp. 205–218, 2005.

[10] S. Fang, T. Lin and K. Lee, “A novel algorithm for multipathfingerprinting in indoor wlan environments,” IEEE Trans. WirelessCommun., 2008.

[11] A. Haeberlen, E. Flannery, A. M. Ladd, A. Rudys, D. S. Wallach,and L. E. Kavraki, “Practical robust localization over large-scale802.11 wireless networks,” in Proc. of ACM MobiCom, pp. 70–84,2004.

[12] T. He, S. Krishnamurthy, T. Luo, L. andYan, K. B., L. Gu,R. Stoleru, G. Zhou, Q. Cao, P. Vicaire, J. A. Stankovic, T. F.Abdelzaher, and J. Hui, “VigilNet: An Integrated Sensor NetworkSystem for Energy-Efficient Surveillance.,” in ACM Transactions onSensor Networks, 2006.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. XX, NO. X, XXXX 2012 11

Kaishun Wu is currently a research assistantprofessor in Fok Ying Tung Graduate School withthe Hong Kong University of Science and Tech-nology. He received the Ph.D. degree in com-puter science and engineering from Hong KongUinversity of Science and Technology in 2011.His research interests include wireless commu-nication, mobile computing, wireless sensor net-works and data center networks.

Jiang Xiao is currently a first year Ph.D studentin Hong Kong University of Science and Tech-nology. Her research interests mainly focusedon wireless indoor localization systems, wirelesssensor networks, and data center networks.

Youwen Yi received his B.Sc (Hons) degreefrom Harbin Institute of Technology in 2007, andreceive his M.Sc degree from Huazhong Univer-sity of Science and Technology in 2009. From2009 and 2011, he was a research assistant atHong Kong University of Science and Technolo-gy . Currently, he is a senior research engineerat Huawei Technologies. His research interestsinclude real-time indoor localization system, andnext generation self-organized network.

Dihu Chen received B. Sc. and M. Phil. de-gree from the Semiconductor Physics of SichuanUniversity in 1986 and 1989, respectively, andhe received Ph.D. degree in solid state electronfrom Department of Electronic Engineering, TheChinese University of Hong Kong in Dec. 2000.Now he is a professor, head of the departmentand ASIC Design Center of Sun Yat-sen Uni-versity. He is currently working on electronicdevices, IC design and design methodology.

Xiaonan Luo is a Professor in the School ofInformation Science and Technology, Director ofthe Computer Application Institute, Sun Yat-senUniversity, China. His research interests includemobile computing, computer graphics and CAD.

Lionel M. Ni is Chair Professor in the Depart-ment of Computer Science and Engineering atthe Hong Kong University of Science and Tech-nology (HKUST). He also serves as the SpecialAssistant to the President of HKUST, Dean ofHKUST Fok Ying Tung Graduate School andVisiting Chair Professor of Shanghai Key Lab ofScalable Computing and Systems at ShanghaiJiao Tong University. A fellow of IEEE, Dr. Ni haschaired over 30 professional conferences.