1.School of Electronics and Information Engineering,Harbin Institute of Technology,Harbin 150080,China;2.School of Communication and Electronic Engineering,Qiqihar University,Qiqihar 161006,China

Radio map updated method based on subscriber locations in indoor WLAN localization

Ying Xia1,2,Zhongzhao Zhang1,*,and Lin Ma1

1.School of Electronics and Information Engineering,Harbin Institute of Technology,Harbin 150080,China;
2.School of Communication and Electronic Engineering,Qiqihar University,Qiqihar 161006,China

With the rapid development of wireless local area network(WLAN)technology,an important target of indoor positioning systems is to improve the positioning accuracy while reducing the online calibration effort to overcome signal time-varying.A novel finge print positioning algorithm,known as the adaptive radio map with updated method based on hidden Markov model(HMM),is proposed.It is shown that by using a collection of user traces that can be cheaply obtained,the proposed algorithm can take advantage of these data to update the labeled calibration data to further improve the position estimation accuracy.This algorithm is a combination of machine learning,information gain theory and finge printing.By collecting data and testing the algorithm in a realistic indoor WLAN environment,the experiment results indicate that,compared with the widely usedKnearest neighbor algorithm, the proposed algorithm can improve the positioning accuracy while greatly reduce the calibration effort.

subscriber location,wireless local area network (WLAN),positioning accuracy,calibration effort,hidden Markov model(HMM).

1.Introduction

The emergence of wireless technology and mobile computing devices has promoted business development of various location estimation systems.In recent years,many researchers focus on offeringperfect services by taking fully advantageof the mobility of the devices.Therefore,the location based services(LBS)have become a hot topic of mobile computing research area.In general,the mobile set needs to locate itself before enjoying the service based on location.It is well known that the existing global positioning system(GPS)[1]may provide a simple and effective solution for such applications in the outdoor environment. In spite of the fact of GPS playing a vital role nowadays, limitations still exist.The main shortage is the invalidity to perform positioning inside the buildings due to the signal blockage in the indoor environment.Thus how to determine the exact location of a mobile device in buildings is still a hot issue to be solved.

Since the widespread deployment of IEEE 802.11 wireless local area network(WLAN)infrastructure has been provided ubiquitous coverage in indoor crowded settings, such as offic building,market,hospital and school. As a result,the network provides fundamental hardware facilities for the developmentof indoorlocation estimation based on radio frequency(RF)technique.Many RF-based systemsinfertheunknownlocationofmobiledevicebythe signal strength received from the detectable access points (AP)in the surrounding environment.They usually work in two steps[2–4]:offlin training phase and online localization phase.In the offlin phase,the area of interest is divided into several grid nodes,meanwhile the so-called referencepointcanbeset asthecenterofgrid.Thenaradio map is established through tabulating the signal strength valuesreceivedat the referencepointsfromthe APs.These values comprise of the physical coordinates and its corresponding signal intensity,which is compiled into a fingeprint database for the online mapping.In the online phase, the real-time signal strength samples received by the mobile terminal from detectable APs are used to lookup the pre-stored fingerprin database to estimate the current location.

However,location estimation is still a challenging problem because of the very complex RF signal propagation. Subjected to the impact of human moving,wall,doors and windows state,layout changes,and radio interference,the signal propagation suffers from severe multi-path fading effects in an indoor environment.As a result,a signalstrength sample measured at a f xed location may greatly deviate from those stored ones in the radio map[5,6].As for most of the indoor positioning algorithms based on fin gerprinting,the constructed radio map in the offlin phaseis viewed as static database and to be used for the later position estimation algorithm[7–10].This approach simplifie the fingerprin database collection as well as the timevarying correction problem.At the same time,it restricts the positioning accuracy improvement and robustness of the system.For the purpose of conquering the received signal variability caused by dynamic environment,several adaptive localization algorithms have been proposed in some literatures[11–15].Yin et al.established a dynamic fingerprin database through constructing the nonliner signal regression model between the reference point and target locations.It is based on the locally linearization assumption.This method is limited effective on constructing a discrete grid style fingerprin database.Furthermore, it is quite difficul to obtain the signal function between reference points and test points.The correlation of signal variation within finit space was used in[14].It calculated the related signal strength value on the basis of real-time signal strength collected from reference beacons.However this method required additional hardware for the complicated arrangement of reference beacons.Other methods attempt to overcome the variations of signal strength with dense reference point distribution,and the accuracy can be guaranteed only when the additional reference points are distributed,which increases the cost of offlin labor effort and online computational time of localization.

Subjectedto the static characteristic limitation of the established radio map in fingerprintin algorithm,the accurate location estimation is still a difficul task.We propose a hidden Markov model(HMM)based algorithm that exploits the user movement trajectory to offset positioning deviations.The user acts as a volunteer to collect realtime data,and the traces record the sequence collection of signal strength during a user’s movement.While the received signal strength data are recorded without any physical location,as a result,these unlabeled samples can be collected easily and cheaply without the labeling process. Using learningalgorithmto extractlocation informationof the unlabeled data,our method provides a way to update the offlin labeled calibration data.Consequently,manual effort can be reduced substantially while in a large-scale indoor environment.The proposed positioning algorithm is still valid.Extensive experiments in results show that our algorithm is effectively reducing the calibration effort under the premise of improving the positioning accuracy.

2.Radio map adaptation based on users moving trajectory

As discussed above,the static characteristic of the radio map collected offlin results in limited accuracy in fingerprint-base positioningapproaches.Forexample,the accuracy of classical RADAR system is about 50%probability within three meters[16].And the LANDMARC systems utilizes dense reference point deployment to mitigate the effects caused by the fluctuatio in radio frequency signal strength,while,it is time-consuming and laborconsuming.To make up with the accuracy loss caused by signal propagation property,we proposed a algorithm that combines the collected radio map offlin with the realtime data sampled by mobile users.Specificall,it uses an HMM to model user traces,and calibrate the radio map through the unlabeled real-time received signal strength (RSS).

2.1 Modeling user unlabeled traces using HMM

HMM is a well-known statistical model.It describes a hidden Markov process containing the hidden parameter.It can be used to speech recognition,natural language processingandbiologicalinformation.Intheupdatingprocess ofradiomapbasedonuser’s trajectory,theunderlyingprocess is a user’s sequential moving traces,where the user’s locations are the hidden internal states,corresponding to theobservablestates arethe signal-strengthmeasurements, which are associated with the unlabeled locations.Thus, considering the RSS values as observable states,the corresponding locations as hidden states,it is possible to achieve the hidden physical position from its observable state through optimizing HMM parameters.Furthermore, the fingerprin data can be updated and maintained effectivelyby meansof using the abovestates.What’s more,the data of reference points can be calibrated to overcome signal fluctuatio caused by environmentalvariation,or serve as new reference points in the fingerprint

In our algorithm,the HMM based on user trajectory is define as〈V,L,A,B,π〉,which includs two state sets and three probability matrix,respectively.

V:Received signal space,V = {v1,v2,...,vn}, vi(1≤i≤n)represents a signal-strength measurement received from the ith access point at reference point.The strength value ranges from–100 dBm to 0 dBm.

L:Location space,L={l1,l2,...,lM},lj(1≤j≤M)represents a reference point labeled physical coordinates in the localization area.

A:Location state transition matrix,A = {aij= P(qt+1=lj|qt=li)}(1≤ i,j≤ M),aijrepresents the transferred probability from liat t period to ljat the next period

B:Fingerprint database,B={bi(k)=P(vk|li)} (vk∈ V,li∈ L),bi(k)represents the conditional probability which gives the likelihood of receiving signalstrength measurement vkin the location liat t period,

π:Initial location state distribution,π=P(li)(li∈L) it represents the prior information where a user may be.

Among the HMM,onlyA,Bandπare the adjustable parameters,and can be represented in term ofθ=(A,B,π).Given an observed user trace and the parameterθ,themostprobablephysicallocationswhicharea sequence of a user’s location changes can be inferred from the observed signal value sequences in the signal space.

2.2Optimizing HMM parameter using expectation maximization algorithm

For the purpose of optimizing the constructed HMM,we utilize a expectation maximization(EM)algorithm to train the model using unlabeled trace and RSS state.

Given a set of tracesT,they are unlabeled data because the signal strength collected by the mobile user as volunteer is recorded without any position label.The EM algorithm is used to adjust the parameterθ=(A,B,π) through iterative computations to finθ?,which can meet the equationcan be calculated as follows:

wheret=(v1,v2,...,vnt)is a received signal strength of lengthnt,andq=(l1,l2,...,lnt)is the corresponding possible location sequence.The probability ofTis the result of the probability of each individual trace,which is a weighted summation over all possible hidden location state sequences under a givenθ.As given in(1),is the probability ofqbeing a user’s true sequence of location changes andP(t|q,θ)is the probability of tracetwhen the user’s movement isq.The two probability values can be expanded as

Thus we can compute the former one fromAandπ,the later fromBrespectively.

The EM algorithm is an iterative process through two steps to maximize theQfunction,which is define as

whereθkis the model parameteracquiredafter thekth iteration.In the iterative process,the sequence of parametersθ0,θ1,...,θ?represents the initial parameterθ0,middle parameter and terminal parameterθ?.The EM algorithm guarantees thatP(T|θk+1)≥P(T|θk),and the parameter convergestoθ?.Consequently,beginning from an initially radio mapB0,the algorithm adjust its parameters iteratively to seek out the best locations of unlabeled traces, improving fingerprin database.If a new radio mapB?is obtained,it can be used to replace the initial oneB0for the later position estimation.

3.Positioning algorithm based on HMM

3.1Data collection in the offlin phase

It is well known that the WLAN localization system based on fingerprin algorithm,such as the famousKnearest neighbor algorithm,consists of two steps:offlin and online.The task of the offlin phase is to construct a radio map for mapping the received signal strength value from a f xed location with its position coordinates.Specificall, the interesting localization area is partitioned into several uniform grids and define the center of grids as reference points(RP).We denoteLg(1≤g≤M)as the locations of RP,Mis the number of total RPs.Then,the RSSs from each AP atLgcan be expressed asRSSi,j∈RM×D, a high dimension dataset withi=1,2,...,M,j=1,2,...,D,Dis the total number of AP.There are two pieces of information stored for each location,as shown in Table 1,physical coordinates and RSS sample values from all the APs.

Table 1 RSS recorded in radio map dBm

3.2AP selection based on information gain theory

With the AP dense deployment environment,it is necessary to exploit a selection criterion to identify the most important APs for HMM updating.Therefore,we select part ofAPs toupdateforonlinepositioningestimationbasedon information gain(InfoGain)method proposed in[17,18]. InfoGain AP selection evaluates the worth of each AP in terms of its discriminative power.Specificall,it is measuredbytheinformationgainwhichcalculatedastheequation in entropy as follows:whereLrepresentsthe coordinateof the referencelocation andH(L)is the entropy of the RP location whenAPj’s value is unknown.H(L)=log2Mis a constant scale ifLfollows a uniform distribution.H(L|APj)is the conditional entropy of the RP location given theAPj’s value, which is equal to

wherevis one possible value of signal strength fromAPjandthesummationistakenoverallpossiblevaluesofAPj. By means of the Bayes’rule,the posterior probability valuesP(Lg|APj=v)can be calculated as

For eachAPj,we calculate the information gain using (6)and rank them in a descending order,then the top APs are selected to participate in the following steps.

3.3Radio map updating based on HMM

According to the locations of the user’s trajectory online stage,we update the radio map based on HMM using the observed signal value sequences in the signal space.More specificall,in the radio map updating process,we collect real-time signals in only part of positioning area for experiments to analyze its influenc on positioning performance.There are total 120 RPs in the experimentalregion, as shown in Fig.1,which are labeled as(l1,l2,...,l120). To simplify the HMM model,the user movement is subjected to move straightforward to nearby locations in consecutive time steps,as the arrow direction shown in the Fig.1.

Fig.1 Experimental test-bed of radio map updating algorithm

We firstl used 40 unlabeled traces(t1,t2,...,t40)for updating,which are sequences of signal strength measurements recording a user’s movements in the environment. The trace of user has no location label assigned when recorded,it is define as a sequence of observed samples (o1,o2,...,o40).Moreover,we consider that the user can only move along three specifie physical trajectories,such asq1,q2,q3intheexperimental.Therefore,thedetailedparameters of HMM model can be described as

(i)L={l1,l2,...,l120},lg(1≤g≤120)represents the 120 reference point locations.

(ii)V={v1,v2,...,v101}represents the received signal strength value at a f xed location from some APs.

(iii)Location state transition matrixA={aij=P(qt+1=lj|qt=li)}(1≤i,j≤120).aijrepresents the transferred probability fromliattperiod toljat the next period,which satisfie the equationaij≤1).

(iv)Signal strength value distribution matrixB={bg(e)=P(ot=ve|qt=lg)},wherebg(e)represents the conditional probability of receiving RSS value isve(1≤e≤101)atlgposition from some AP duringtperiod.It satisfieso we can construct the matrix through statistical analyzing offlin fingerprin database.

(v)Initial location state distribution matrixπ=Because three initial points are specifie as the following uniform distribution,so that theinitial location distribution probabilities of other points are equal to zero.

3.4Real-time positioning in the online phase

For the online phase,the RSS sampled at the test point is represented asRSSt,j∈R1×D.At first the HMM model is used to update the statistical radio map obtained in the offlin phase.And then,the deterministic weightedKnearest neighbor algorithm is used to calculate the Euclideandistancesbetweenthetest pointandeachRP onthe basis of(7).

whereD?is the number of AP selection result,which is performed in the previous step based on InfoGain criterion.Finally,the mostk(k≥1)RPs with the shortest distancedgare chosen to estimate the locationof the test point by weighting these RPs’location on the basis of (8).

4.Experiment results and analysis

4.1Experimental setup

We performextensiveexperimentsin a realistic WLAN indoor positioning scenario,as shown in Fig.2,including corridors and some rooms in our department.The interesting area for localization is the corridor,which has a dimension of 61.4 m×24.7 m.There are total 27 Linksys WRT54G APs(Cisco,Irvine,California,USA)which are deployedin 2 m height from the ground.The blue symbols show their locations in the whole floo plan.In addition to achieve full coverage of the region,the high density deployment of the access points can guarantee that any position may receive six to 27 sample signals from different APs.Obviously,most of the wireless channelsare non-line of sight and influenc by the existing barrier wall.

Fig.2 Layout of the experimental test-bed

In the interesting area for localization,we divide the whole corridor into several uniform 1 m×1 m grids and denote the center of each grid as a reference point.Altogether,there are 247 RPs in our experiment environment, part of them are marked with red star symbols in above Fig.2.We adopt an ASUS A8F laptop as the terminal node,with API program“HITWLAN v1.0”operating system that runs under Windows XP.An Intel PRO/Wireless 3945ABG wireless card is installed to gather signal samples from nearby APs at two samples per second in the offlin phase.For undetected APs,we set a default value,–95dBm,the minimumdetectableRSS.Inmoredetail,we gathered 400 received signal strengths samples from four different directions at each location of the RP,and each direction is 100.For the presence of RF signals having complex time-varying statistical properties,we treat samples after time diversity learning to construct the radio map. Therefore,the radio map comprised of received signal val-ues with its corresponding physical coordinate is a dataset with 247×29 dimensionality.

In the online phase,we use a feature matching algorithm to estimate the location.The positioning accuracy of widespread adoption algorithms,such as the WKNN, KNN and NN,are illustrated in Fig.3.Obviously,the WKNN algorithm has the attractive performance:as the number ofkincreases from one to six,the accuracy within 2 m increases gradually.As the value ofkis six,the WKNN algorithm can achieve best accuracy of 76.8%,as shown in Fig.3(a).Further increasing its value will result in the system performance deterioration and online positioning computationalaggravation.However,in our experiment,the online test data which are selected at the predefine RP locations will led the NN algorithm to hare the highest accuracy within 1 m.For the only feature,it can notbeconcludedthattheNNis outperformingtheWKNN. The performance of WKNN is stably outperform the others all along ask=6,shown in Fig.3(b).For this reason, in the subsequentexperiments,we setkto six for positioning performanceanalysis using the WKNN algorithm.The Euclidean distance between the true coordinate and the estimated one is deployed as performance of the proposed technique metric which is called a error distance.Positioning accuracy is indicated by the cumulative error distance distribution.

Fig.3 Effect ofkon positioning performances

Given the model initial parameterθ=(A,B,π)and real-time user tracesT=t1,t2...,tT,we can infer the most probable sequence of location changesQ=q1,q2...,qTthrough the decoding algorithm.

4.2Updated algorithm performance analysis based on information entropy gain

Usually,radio map updatingperformancecan be evaluated by degree of the updated data consistent with real-time measurements.On the other hand,it can be directly used to indoor positioning and its performancecan be evaluated through location estimation accuracy.For this purpose,we take the updated data into positioning process and analyzing system performance.

In the offlin phase,we study the relationship between performancewithvariednumberofAPsbasedonthreedifferent AP selection criteria.Fig.4 compares the accuracy within 2 m in Fig.4(a)and the average error distance in Fig.4(b)respectively,using the deterministic weightedKnearest neighbor algorithm over all the locations with respect to three AP selection criteria:InfoGain,Max-mean, and Random.

Fig.4 Effect of AP selection methods on performance

In order to reveal the performanceof the InfoGain criterion,we compare it with Max-mean and Random criteria. Specificall,the InfoGain criterion ranks APs in descending order of their information entropy gain values as de-scribedin Section 3.2.While the Max-meanoneranks APs in descending order of their average signal strength values, and the random criterion randomly selects a few APs regardless of their signal strength values.

From the performance curves,we can fin that as the numberof APs increases,the accuracywithin 2 m of using theInfoGainmethodincreasesfasterthantheothertwocriteria.In other words,the InfoGain selection criterion measure discriminantofeach AP is moreprecise.As thesubset of APs is equal to 16,the InfoGain criterion can achieve the best accuracy of 76.8%within 2 m,while the accuracy of the Max-mean is 67.8%as the APs increase to 16.On the contrary,the Random one executes worst resulting in randomly AP selection.This shows that the InfoGain criterionhas the advantageof usingthe fewest APs to achieve the same accuracy,which in turn reduces the online positioning computational cost.

Furthermore,we know that the localization area with increasing number of APs does not equal to positioning accuracyimprovement.In fact,throughthe InfoGaincriteria, once the optimal subset of APs has been selected to participate in positioning process,additional usage of the other APs can introduce interferences which bring positioning accuracy decreasing.To this end,in the rest of our experiments,weset thenumberofbeingupdatedAPto16forour experiment analysis.Fig.5 shows the improvement in accuracy using unlabeled traces.The four curves in Fig.5(a) show the comparisonon performancechanges betweenthe originaland the updatedalgorithmperformance,which are based on different AP selection algorithms.It is clear that the algorithm based on InfoGain criterion outperforms the one based on Max-mean criterion.There are about 5%improvement within 2 m after using 40 unlabeled traces to updating fingerprintin dataset.To investigate the usage of unlabeled traces,we also varied the number of traces used in training the HMM.As the number of traces increases, from 0 to 100,the accuracy goes up monotonously,as shown in Fig.5(b).Specificall,when the numberof traces used is zero,the original accuracy within 2 m is about 76.7%.Improvement is about 0.7%when 20 traces are used and 5.9%using 100 traces.However,the accuracy cannot be improved all along through increasing the number of traces.Thatis to say,positioningimprovementtends to stabilize as more traces are used.To our experimental scenario,the optimal trace is 60.Moreover,we illustrate the improvements within different error distances as the number of trace varied in Fig.6 and confir that the satisfie performance improvement can be achieved by using 60 unlabeled traces.

Fig.5 Accuracy improvement after updating

Fig.6 Accuracy improvement using different traces

5.Conclusions

Inthis paper,the effectof reducingthe calibrationefforton estimation accuracy caused by radio propagation fluctua tion in the widely spread WLAN indoor localization based on fingerprintin technique is studied.More specificall, tediousandrepeatedmanualeffortis requiredforthe dense reference point deployment and late calibration in order to overcome time-varying characteristics of the wireless signal.When additional user traces are available,the HMM-based algorithm can take advantage of these visible signal data to update the labeled calibration data to further improve position estimation accuracy.In addition,we use an InfoGain AP selection technique to enhance the efficien y of the proposed algorithm.The calibration of radio mapis achieved by decoding the hidden location states from the HMM using EM algorithm.Experimental results show that our proposed method can obtain nearly 6%accuracy improvement within 2 m(using only 60 unlabeled traces to updating the radio map)and meanwhile reduce the calibrationeffortgreatly.In thefuture,we will considerhowto speed up offlin parameters training.Thus,our algorithm can perform in large-scale dynamic environment,and the robustness of our proposed algorithm will be studied.

[1]E.C.L.Chan,G.Baciu,S.C.Mak.Wireless tracking analysis in location fingerprintingProc.of the IEEE International Conference on Wireless and Mobile Computing,2008:214–220.

[2]M.Youssef,A.Agrawala,U.Shankar.WLAN location determination via clustering and probability distributions.Proc.of the 1st IEEE International Conference on Pervasive Computing and Communications,2003:143–150.

[3]C.Feng,W.S.A.Au,S.Valaee,et al.Received-signalstrength-based indoor positioning using compressive sensing.IEEE Trans.on Mobile Computing,2012,11(12):1983–1993.

[4]S.H.Fang,T.N.Lin.Principal component localization in indoor WLAN environments.IEEE Trans.on Mobile Computing,2012,11(1):100–110.

[5]S.H.Fang,T.N.Lin,K.C.Lee.A novel algorithm for multipath fingerprintin in indoor WLAN environments.IEEE Trans.on Wireless Communications,2008,7(9):3579–3588.

[6]S.Mazuelas,A.Bahillo,R.M.Lorenzo,et al.Robust indoor positioning provided by real-time RSSI values in unmodifie WLAN networks.IEEE Journal of Selected Topics in Signal Processing,2009,3(5):821–831.

[7]M.Youssef,A.Agrawala.Handling samples correlation in the Horus system.Proc.ofthe23rd Annual JointConference ofthe IEEE Computer and Communications Societies,2004:1023–1031.

[8]S.P.Kuo,Y.C.Tseng.A scrambling method for fingeprint positioning based on temporal diversity and spatial dependency.IEEE Trans.on Knowledge and Data Engineering, 2008,20(5):678–684.

[9]L.Wirola,T.A.Laine,J.Syrj¨arinne.Mass-market requirements for indoor positioning and indoor navigation.Proc.of the International Conference on Indoor Positioning and Indoor Navigation,2010:1–7.

[10]L.Ma,C.Zhou,D.Qin,et al.Green wireless local area network received signal strength dimensionality reduction and indoor localization based on fingerprin algorithm.International Journal of Communication Systems,2013,27(12).

[11]A.Kushki,K.N.Plataniotis,A.N.Venetsanopoulos.Intelligent dynamic radio tracking in indoor wireless local area networks.IEEE Trans.on Mobile Computing,2010,9(3):405–419.

[12]J.Yin,Q.Yang,M.N.Lionel.Learning adaptive temporal radio maps for signal strength based location estimation.IEEE Trans.on Mobile Computing,2008,7(7):869–883.

[13]Y.M.Lin,H.Y.Luo,J.T.Li,et al.Dynamic radio map based particle filte for indoor wireless localization.Journal of Computer Research and Development,2011,48(1):139–146.(in Chinese)

[14]P.Krishnan,A.S.Krishnakumar,W.H.Ju,et al.A system for LEASE:location estimation assisted by stationary emitters for indoor RF wireless networks.Proc.of the 23rd Annual Joint Conference of the IEEE Computer and Communications Societies,2004:1001–1011.

[15]L.M.Ni,Y.Liu,Y.C.Lau,et al.LANDMARC:indoor location sensing using active RFID.Proc.of the 1st IEEE International Conference on Pervasive Computing and Communications,2003:407–415.

[16]P.Bahl,V.N.Padmanabhan.RADAR:an in-building RF-based user location and tracking system.Proc.of the International Conference on Computer Communications,2000:775–784.

[17]Y.Q.Chen,Q.Yang,J.Yin,et al.Power-efficien AP selection for indoor location estimation.IEEE Trans.on Knowledge and Data Engineering,2006,18(7):877–888.

[18]N.Alsindi,Z.Chaloupka,J.Aweya.Entropy-based location fingerprintin for WLAN systems.Proc.of the International Conference on Indoor Positioning and Indoor Navigation, 2012:1–7.

Biographies

Ying Xiawas born in 1973.She received her B.S. degree in automation and control from Qiqihar University,in 1996,and M.S.degree in communication and information system from Tianjin University,in 2007,respectively.She is an associate professor in the School of Communication and Electronic Engineering,Qiqihar University,Qiqihar,P.R.China.Currently,she is aPh.D.candidate in the Communication Research Center of Harbin Institute of Technology,Harbin,P.R.China.She is a member of Heilongjiang Province Institute of Electronics.Her research interests include location-based service for wireless local area networks,wireless communication networks,artificia intelligence and manifold learning.

E-mail:xyingw@hit.edu.cn

Zhongzhao Zhangwas born in 1951.He received his M.S.degree in Communications Engineering from Harbin Institute of Technology,in 1984.Currently he is the director of Communication Research Center of Harbin Institute of Technology and a full professor at School of Electronics and Information Engineering of Harbin Institute of Technology.He is an editorial board member of several domestic prestigious journals.He has conducted various types of researches related to the National Natural Science Foundation of China,the National High Technology Development Program,the National Basic Research Program,the National Science and Technology Major Project and the National Key Project.His research interests include satellite communications,wireless mobile communications,digital signal processing,and GNSS receiver techniques.

E-mail:zzzhang@hit.edu.cn

Lin Mawas born in 1980.He received his B.S., M.S.,and Ph.D.degrees in Electrical Engineering from Harbin Institute of Technology,in 2003,2005, and 2009,respectively.He engaged the postdoctoral program in Computer Science and Technology Postdoctoral Station of Harbin Institute of Technology,from 2009 to 2011.Since 2013,he is a visiting scholar at the Edward S.Rogers Sr.Department of Electrical and Computer Engineering at the University of Toronto, Canada.He is an associate professor in the Communication Research Center of Harbin Institute of Technology,China.He is a member of the IEEE.His research interests include location-based service for wireless local area networks,cognitive radio,and cellular networks.

E-mail:malin@hit.edu.cn

10.1109/JSEE.2015.00131

Manuscript received January 19,2015.

*Corresponding author.

This work was supported by the National Natural Science Foundation of China(61571162)and the Major National Science and Technology Project(2014ZX03004003-005).