Research Article Indoor Localisation Based on GSM Signals: Multistorey Building Study

Transcription

1 Moble Informaton Systems Volume 26, Artcle ID , 7 pages Research Artcle Indoor Localsaton Based on GSM Sgnals: Multstorey Buldng Study RafaB Górak, Marcn Luckner, MchaB Okulewcz, Joanna Porter-Soberaj, and Potr Wawrzynak 2 Faculty of Mathematcs and Informaton Scence, Warsaw Unversty of Technology, Koszykowa 75, -662 Warsaw, Poland 2 Orange Labs Poland, Obrzeżna 7, 2-69 Warsaw, Poland Correspondence should be addressed to Rafał Górak; r.gorak@mn.pw.edu.pl Receved 8 January 26; Accepted 28 March 26 Academc Edtor: Robert Pché Copyrght 26 Rafał Górak et al. Ths s an open access artcle dstrbuted under the Creatve Commons Attrbuton Lcense, whch permts unrestrcted use, dstrbuton, and reproducton n any medum, provded the orgnal work s properly cted. Among the accurate ndoor localsaton systems that are usng WF, Bluetooth, or nfrared technologes, the ones that are based on the GSM rely on a stable external nfrastructure that can be used even n an emergency. Ths paper presents an accurate GSM ndoor localsaton system that acheves a medan error of 4.39 metres n horzontal coordnates and up to 64 percent accuracy n floor predcton (for 84 percent of cases the floor predcton s mstaken by not more than a sngle floor). The test and reference measurements were made nsde a sx-floor academc buldng, wth an rregular shape, whose dmensons are around 5 metres by 7 metres. The localsaton algorthm uses GSM sgnal readngs fromthe 7 strongestcells avalable n the GSM standard (or fewer, f fewer than 7 are avalable). We estmate the locaton by a three-step method. Frstly, we propose a pont localsaton soluton (.e., localsaton based on only one measurement). Then, by applyng the central tendency flters and the Multlayer Perceptron, we buld a localsaton system that uses a sequence of estmatons of current and past locatons. We also dscuss major accuracy factors such as the number of observed sgnals or the types of spaces n the buldng.. Introducton Outdoor localsaton s, today, a part of our lfe. However, such useful localsaton methods as the Global Postonng System (GPS) fal nsde buldngs. Popular alternatves use Receved Sgnal Strength (RSS) from wreless networks that are accessble ndoors. Measurng the strength of WF sgnals from multple Access Ponts n varous locatons, we create a map of fngerprnts. In the localsaton process one s poston can be found by comparson of current sgnal strengths wth the created map. Usng WF sgnals nsde buldngs wth many dfferent Access Ponts whose range covers the whole buldng allows a very accurate localsaton soluton to be bult. However, ths method may fal n buldngs wth a poor network nfrastructure. Moreover, n the case of an emergency such as a fre n the buldng, local nfrastructure may be damaged and the localsaton system wll fal. An alternatve s a localsaton system based on an outsde nfrastructure. We present an ndoor localsaton systembasedonglobalsystemformoblecommuncatons (GSM) sgnals. The proposed localsaton system can be mplemented on most Androd moble phones. However, there are exceptons. Our test showed that Samsung Galaxy S III returns lmted data and cannot be used. We tested the system n a sx-floor academc buldng for three-dmensonal localsaton of objects. The paper presents the followng results: a horzontal localsaton soluton wth the medan error less than 4.39 metres and a floor detecton soluton wth an accuracy of 64 percent. The remander of ths work s organsed as follows: Secton 2 descrbes the data and the data acquston

2 2 Moble Informaton Systems process. Secton 3 presents the two-stage localsaton process. Secton 4 presents the test results. The dscusson s presented n Secton 5, where major accuracy factors are analysed, and Secton 6, where our results are compared wth other algorthms. Ths s followed by the conclusons n Secton Data 2.. Data Descrpton. The data for the statstcal models were collected usng three dfferent moble phones, Sony Ercsson E5,SonyErcssonMT,andHTCDesre,allofthem runnng Androd OS 2. (or newer). The measurements were taken wth the phones held horzontally by hand at around metre above the floor. All moble phones were attached to a GSM (2G) moble network durng the experment. It should be noted that n parallel wth collectng the data on moble phones, the Gateway Moble Locaton Centre- (GMLC-) based locaton of the termnal was estmated, whch could result n more frequent cell reselecton requests sent to a moble staton (compared wth a termnal operatng n a purely dle mode of operaton). Ths results n a more complete representaton of possble servng cells n a gven locaton, although none of our models takes drect advantage of that fact, relyng only on vectors of Receved Sgnal Strengths. The data were gathered on all of the publcly accessble areas of a sx-floor academc buldng (ncludng the ground floor). The buldng has an rregular shape, ts outer dmensons are around 5 by 7 metres, and ts heght s 24 metres. A 3D sketch of the buldng s shown n Fgure. For moredetaledplansofthebuldngyoumayvsttheportal for the ongong ndoor localsaton project ( For the purpose of tranng and testng the statstcal models for localsaton predcton, we gathered the data n three ndependent seres of measurements (each seres took place n a dfferent week). In order to have accurate coordnate nformaton nsde the buldng, we defned a.75-by-.75-metre grd and assgned for each pont a unque dentfer (denoted by POI). All measurements were taken n the ponts of the defned grd and they were labelled by the correspondng POIs. A sngle measurement (fngerprnt) s a real valued vector F=(rss, rss 2,...,rss N ),whererss k s the RSSI from the kth BTS, typcally rangng from 3 dbm to 5 dbm. The dmenson N of the space of the fngerprnts s the number of all BTSs that we decded to nclude when creatngthelocalsatonmodel.inourcase,thssthenumber of all BTSs that were observed nsde the buldng, that s, N=39. When there was no sgnal from the kth BTS, we set rss k = 3. The data were gathered by three persons and each seres of measurements took 9 days on average. Each seres of measurements was taken n around 2 POIs. To make the fnal results more ndependent from spatotemporal RSSI fluctuatons, at each POI the measurement was taken 4 tmes.thsresultednatotalnumberofca.48fngerprnts. The frst and the second seres were taken n the same set of POIs, ordered n general n a.5-by-.5-metre grd (denoted GRID,2 ), whle the thrd seres used for the test was Fgure : Overall presentaton of the acqured measurements wthn a 3D sketch of the buldng. The green areas represent the rooms and sectons of the buldng where the measurements were taken and thevertcalbarsdenotethelocatonswherethemeasurementswere taken. Table : Descrptve statstcs of the gathered data for each floor. Floor #BTS Q Q 2 Q 3 Samples s x [m] s y [m] #area % % % % % % taken n another set of POIs whch formed another.5-by-.5- metre grd (denoted GRID 3 ) shfted by.75 metres offset n each drecton from GRID,2.Theexceptonsfromthe.5-by-.5-metre grd were the large lecture halls where a 3.-by-3.- metre grd was used nstead. In the remander of the paper we shalldenotethefrst,thesecond,andthethrdseresbys, S 2,andS 3,respectvely.Obvously,S R 39 for {,2,3}. Feature vectors (elements of S )werelabelledby () the value of the x and y coordnates (n metres), () the floor number, () one of the four drectons n whch the phone was orented (parallel to one of the horzontal axes); t should be noted that the orentaton nformaton has not been passed to the predcton model but was used n the generaton of artfcal paths. In the whole buldng we observed 39 dfferent BTS dentfers. RSSI values for up to 7 BTSs were regstered n a sngle measurement, whch corresponds to the GSM standard. In partcular, accordng to the Rado Resource Control Protocol [], a regular Moble Electroncs s allowed to report the sgnal strength for servng cell and up to 6 neghbourng cells n one measurement report. Table presents selected statstcs on each floor of the buldng for the number of BTSs, relatve sample sze, standard devaton n horzontal drectons, and the number of arbtrarly chosen areas (rooms, dstnctve sectons of the corrdors, halls, etc.) Despte the approprate standards, we observed that moble phones used n the experment reported RSSI values that are not allowed by the standard. In partcular, RSSI

3 Moble Informaton Systems Fgure 2: Dstrbuton of observed RSSI values. levels rangng from 3 dbm to were observed, whle 3GPP specfcatons defne the maxmum value as 5 dbm (whch s to be nterpreted as 5 dbm or greater ). Such measurements form.4 percent of the data set. Moreover, unknown RSSI values were reported by the moble termnal n.7 percent of the measurements. To sum up, after data set valdaton,atotalnumberof683258regsteredrssivalues forgsmbtsswereusedforthepurposeofthswork.a hstogram showng the dstrbuton of the vald RSSI values regstered n our database s shown n Fgure 2. The measurements scope was focused on the halls and corrdors of the buldng, wth the addton of several lecture rooms and computer laboratores, as these are the most often vsted locatons wthn the buldng. Fgure presents an overall vew of the buldng wth the markers denotng the ponts of GRID,2 and GRID 3. The data descrbed n ths secton was also used n the analyss presented n [2, 3], ncludng the data from the WF Access Ponts (whch are not dscussed n ths work) Artfcal Path Generaton. Although the basc predcton models process only a sngle fngerprnt, much better results (especally n terms of determnng the floor) could be acheved wth a multpont method (.e., takng nto account the past measurements). In each seres of measurements, ateverypoiwetookfngerprntsneachofthefour man drectons parallel wth the coordnate axes of the buldng. We used one of the fngerprnts from a gven pont to construct artfcal paths between two randomly chosen ponts from the grd of defned and measured ponts. We generated 5 paths for the S and S 2 measurement seres and paths for the S 3 -testng measurement seres. The paths were chosen by traversng the graph spanned by the ponts from the gven seres (GRID,2 or GRID 3 ). The neghbourhood of the pont was a set of all the ponts at a dstance of exactly.5 metres. If n a gven drecton there was no such pont, a pont at a 3.-metre dstance (large lecture halls) and subsequently a pont at a.75-metre dstance (narrow corrdors) were chosen as ts neghbours. The topology of the buldng was not taken nto account durng the constructon of the neghbourhood, except for nformaton about the postons of the starwells, whch was used for changng the floor. The paths for the S seres ( =, 2, 3) were generated wth thehelpofrsoftwareusngthefollowngalgorthm: () Choose at random a start pont from the S seres. (2)ChooseatrandomanendpontfromtheS seres. (3) Fnd the shortest path n the graph of ponts (a change of floor s possble only near the man starcase). (4) Choose a fngerprnt gathered n the drecton n whch the path s traversed, havng RSSI for at least two BTSs. In the presented generator, f the start and end ponts are from two dfferent floors, the algorthm searches for the shortest path from the start pont to the starcase, then nserts the predefned ponts between the start and end pont floors, and then searches for the shortest path from the starcase to the end pont. The shortest path was found usng the greedy A algorthm(cf.[4]).theneghbourofthecurrentpont,whch mnmsed the dstance to the end pont (or starcase), was chosen as the next pont on the path. 3. Methodology Our task was the detecton of the current locaton of the tracked object. The locaton s descrbed as a trple p = (x, y, f),wherex and y are the horzontal coordnates of the tracked object and f sthenumberofthecurrentfloor.the coordnates x and y belong to R, whle f belongs to Z +.The localsaton system took as the nput the seres of fngerprnts (F,F 2,...,F k ),wheref R 39. The fngerprnts were taken n consecutve ponts of the movement path where the kth coordnate s the last fngerprnt. Fgure 3 presents the schema of the localsaton process that goes separately for each coordnate of p (.e., for p =x, p 2 =y,andp 3 =f). The process s herarchcal and has three steps for each jth coordnate (j {,2,3}). Step. The estmaton of the coordnate where the fngerprnt F was taken for every k.thswayweobtanedk estmatons (so-called pont estmatons) { p j } k that represent the movement path of the termnal. Ths process s descrbed n detal n Secton 3.. Secton 4. descrbes several machne learnng methods and we choose the best performng one. Fnally Secton 5 provdes an analyss of the major accuracy factors for the pont estmaton method, as ths s the man component of the whole localsaton process. Step 2. Two estmatons are p j t and pj w of the jth coordnate of the poston where the last fngerprnt F k was taken. Ths stepsbasedontheresultsfromstepone.theobtanedpont p t = (p t,p2 t,p3 t ) s the average result of the estmatons { p,..., p k }.Fortheselectedk, such average estmaton of the poston where F k was taken can brng better results than

4 4 Moble Informaton Systems F Calculate weghts (w,...,w k ) Estmaton n pont p p F 2 Weghted tendency flter p w Estmaton n pont p 2 p 2 Tendency flter p t MLP p m F k Estmaton n pont p k p k Estmaton for sngle fngerprnt Estmaton for fngerprnts sequence MLP aggregaton Fgure 3: Localsaton schema for pont p. The scheme presents three steps of estmaton that are done separately for p =x, p 2 =y,and p 3 =f. the estmaton based on a sngle fngerprnt, descrbed n Step. Obvously, the parameter k (the path length) cannot betoolarge.toreducethenfluenceofnotwellestmated ponts, we can make a second estmaton p w =(x w,y w,f w ) that calculates the central tendency wth weghted sngle estmatons. The weght should be selected to stress the nfluence of well estmated ponts and s based on the number of reported BTSs for the fngerprnt F (.e., a number between and7),whchwefoundtobeanmportantestmaton relablty factor. The detals are presented n Secton 3.2. Step 3. The fnal estmaton for a sngle fngerprnt s the functon F k p j k that, for the last observed fngerprnt, returns the estmaton of the jth coordnate (j {,2,3})ofthe pont p k where the fngerprnt F k wastaken.itsbasedon the results from Steps and 2. In ths stage, we had the three estmatons of the poston where F k was taken: p k, p t,and p w. We may assume that the estmaton p w wll brng the best results among all estmatons provded that the weghts were well selected. However, we can try to mprove our result usng a method that aggregates all estmatons. For that the Multlayer Perceptron can be used. The estmatons consttute an nput for the Multlayer Perceptron that calculates the fnal estmaton p m =(x m,y m,f m ) of the poston where the fngerprnt F k was taken. The detals are descrbed n Secton 4.3. The localsaton schema (Fgure 3) s the same for all coordnates. However, there are some dfferences between estmaton for the dscrete floor f and the estmaton of the contnuous coordnates x and y. The man dfference les n the measure of central tendency that s used. For the floor estmaton, a mode flter s used, whle for the estmaton of coordnates an average flter s appled. Detals about all components of the localsaton process are gven n the followng sectons. 3.. Estmaton for a Sngle Fngerprnt. Three functons F p j where j {,2,3} that estmate the coordnates of the localsaton p on the bass of a sngle fngerprnt F are mplemented as an ensemble of decson trees. The ensemble contans multple estmaton or predcton trees created on the bass of varous fragments of the learnng set. A combnaton of the results obtaned by the trees produces the fnal, aggregated estmator or predcton. We compared several boostng algorthms to estmate the localsaton separately for the floor and the coordnates. Our test ncluded AdaBoost, Baggng, and Least Squares Boostng (LSBoost). AdaBoost s a classfcaton method [5], Baggng can be used for classfcaton or regresson [6], and LSBoost s a regresson method [7]. The floor estmaton can be performed as classfcaton or regresson task. The estmaton of coordnates s a regresson task. AdaBoost was mplemented as the AdaBoostM2 algorthm,whereweghtedpseudolossscalculatedforn observatons and K classes: ε t = 2 N d (t) n= k=y n n,k ( h t (x n,y n )+h t (x n,k)), () where h t (x n,k) s the confdence of predcton, d (t) n,k are observaton weghts, and y n s the true class label. Pseudoloss s a measure of the classfcaton accuracy from any classfer n an ensemble.

5 Moble Informaton Systems 5 Baggng bags a weak classfer such as a decson or regresson tree on a data set, generates many bootstrap replcas of ths data set, and grows decson trees on these replcas. To fndthepredctedresponseofatranedensemble,thealgorthm takes an average of predctons from ndvdual trees. LPBoost performs multclass classfcaton by attemptng to maxmse the mnmal dfference between the predcted softclassfcatonscoreforthetrueclassandthelargestscore for the false classes n the tranng set. Ths operaton should mprove generalsaton ablty [8] Estmaton for the Fngerprnts Sequence. Ths part descrbes Step 2 of our method; that s, we fnd two estmatons of the poston where the last fngerprnt F k was taken. We denote them by p t =(x t,y t,f t ) and p w =(x w,y w,f w ) based on the seres of estmated ponts { p =( x, y, f )} k n the prevous secton Mode Flter. The mode flter calculates the most popular result among the floor estmatons for a sngle fngerprnt: k f t = mode ( f ). (2) = j= The flter s parametersed by the coeffcent k that defnes the sze of consdered hstorcal estmatons. The dea behnd that formula s that a sngle estmaton of another floor n a sequence of equal estmaton s a mstake, rather than an expected movement. To avod a strong nfluence of the estmatons for the farthest ponts, estmaton s repeated tmes, where s the poston of the estmaton n the estmatons sequence and k sthepostonofthelastestmatedpont. Equaton (2) assumes that all the ponts are estmated wth the same precson. The followng formula ntroduces an addtonal coeffcent w that symbolses the qualty of the estmaton: k w f w = mode ( f ). (3) = j= In the modfed formula, each estmaton s replcated addtonally w tmes. The value of w should be set expermentally as a measure correlated wth the estmaton qualty Average Flter. The average flter has the same role as the mode flter but works wth contnuous data. The x and y coordnates can be estmated usng the followng formula: x t = k = x, (4) k k = y t = k = y. (5) k k = The element (k k = ) normalses the average, whch s calculated wth a decreasng nfluence of the estmated ponts. Smlarly to the mode flter, we ntroduce the second form of the flter that ncludes knowledge about the qualty of the estmaton: x k = x w = w k k = k = w, (6) k = y w = w k k = k = w. (7) The weghted average uses weghts w that estmate the qualty of the x and y estmatons. The normalsaton factor s (k k = k = w ). Unlke the commonly used complex methods, such as Kalman s flter [9], the proposed flters are qute smple. Theflterscanbeusedforalldatawthcalculatedweghts. Moreover,themethodcanbeusedondatawthoutdvson nto the learnng set and the testng set f the parameter k s already fxed Aggregaton by a Multlayer Perceptron. After the estmaton for the fngerprnts sequence we had the p k, p t,andp w estmatons for the pont p where the last fngerprnt F k was taken. Although we assumed that, for the correctly selected w weghts, p w wouldbethebestestmaton,therestofthe estmatons could gve some addtonal knowledge. To use all that avalable knowledge, we decded to aggregate all the estmatons. For that task, we used a Multlayer Perceptron. We created two ndependent MLP models. The frst model detected the floor and the second model detected the horzontal coordnates. All nputs and outputs of the frst network were bnary. Therefore, all the features were represented by a set of neurons andeachneurondefnedoneofthepossblevaluesofthe feature. Only one neuron from the set could be actve at any one tme. If the number of floors s gven by the number n f,the output layer wll have n f neurons and only one of them s actve. The same s wth nput features. The network consders the estmated floors f,...,f k,theweghtsw,...,w k,the mode f t (2), and the weghted mode f w (3). All features, except the weghts, descrbe floors; therefore the total number of nput neurons s n f k+7 k+n f +n f =n f (k+2) +7k. (8) The model for coordnates s much smpler. All the nputs and the output are contnuous. The nput neurons represent estmated values x,...,x k (or y,...,y k )andweghts w,...,w k, x w, x t (or y w, y t ), so the nput layer has 2k + 2 neurons. The output result s one contnuous value x m or y m dependng on the nput. The locaton estmated by MLP should gve better results than a pure tendency flter. However, the created network wll be ftted to the problem descrbed by the learnng set. In the learnng process, two sets must be created. A learnng set defnes a network structure and a valdaton set controls learnng process and prevents overlearnng. Although such y

6 6 Moble Informaton Systems Method Table 2: Floor detecton. Results for the testng seres S 3. Accuracy [%] Average error [floor] AdaBoost 33.5 Baggng Classfcaton Baggng Regresson 43.8 Least Squares Boostng 29.7 dvson allows generalsaton, the created network cannot be used on a testng set wth a dfferent number of floors, whle amodeflterworksonanydata. 4. Tests Ths secton presents the results obtaned n the buldng descrbed n Secton 2. The results of the floor detecton and the results of the coordnates approxmaton are gven separately for each of the three stages: the estmaton for a sngle fngerprnt, the estmaton for the fngerprnts sequence, and the MLP aggregaton. 4.. Estmaton for a Sngle Fngerprnt. Data for the tests were dvded nto the learnng set, the valdaton set, and the testng set created from the frst seres of measurements S, the second seres of measurements S 2,andthethrdseresof measurements S 3, respectvely. Estmatons for the floor and coordnates were made separately usng dfferent algorthms Floor Detecton. We tested four classfcaton boostng algorthms to estmate the current floor of the tracked object. Our test ncluded AdaBoost, Baggng Classfcaton, Baggng Regresson wth dscretsaton of the results, and Least Squares Boostng. Table 2 presents the average error and accuracy for all the methods. The Baggng algorthm for the classfcaton task gave the best results. The obtaned level of errors s rather hgh, but the value of the average error suggests that n most ofthecasesthepredctonerrorsnotbggerthanonefloor Coordnates Approxmaton. Estmaton of horzontal poston s a regresson task. We checked two regresson models to solve ths problem: the Baggng method and the LS Boostng method. The results are gven n Table 3. The better results were obtaned by the Baggng method. The mean error for the Baggng method could be accepted n ndoor localsaton, but gross errors were very frequent. The locaton estmatons of 2 percent of all the test fngerprnts were mstaken by more than 2 m horzontally Estmaton for the Fngerprnts Sequence. Let us consder asequencef,f 2,...,F l of the fngerprnts that were taken along one of the artfcal paths (constructed as descrbed n Secton 2.2). For each fngerprnt F, let us denote a pont p = ( x, y, f ) that estmates the locaton where F was taken. The estmaton s created usng the Baggng method as descrbed n Secton 3. and Secton 4.. Table 3: Horzontal error analyss for testng seres S 3. Mean Medan 8th percentle LS Boostng Baggng The am of ths secton s to descrbe the experments forcomputngthelocatonofthepontwheref was taken on the bass of F k+,f k+2,...,f sequence of readngs. For < k, we wll proceed by takng the sequence F,F,...,F,F 2,F 3,...,F of length k. Below,weanalysedfferent values of k {,2,...,9}, n order to choose one that provdes the best accuracy. The defned sequence of fngerprnts of length k allows us to apply the methods descrbed n Secton Weghts Selecton. We ntroduced formulas (3) and (6) to operate wth sgnals wth a dfferent qualty. In applcaton, t s necessary to fnd an accurate estmaton of the qualty. We expected to have a better accuracy when more BTSs were vsble. We say that a partcular BTS s vsble for a gven fngerprnt when the devce reported a sgnal from that BTS. It s worth recallng that commonly used devces report up to 7 sgnals (usually the strongest ones) and the senstvty s 3 dbm (sgnals that are weaker are not reported). Fgure 4(a) represents the results of the floor classfcaton for a test set of fngerprnts wth respect to the number of BTSs that are vsble. The results are as we expected. For nstance, when 7 BTSs are vsble, the accuracy s up to66percent,whlewthonlyvsblebtstheaccuracy was as low as 34 percent. Ths suggests that when localsaton s based on a hstory of readngs n the next steps of our algorthm, one should gve more mportance to the fngerprnts wth more BTSs that are vsble. Smlarly to the floor estmaton case, we nvestgated the accuracy of the Baggng method wth respect to the densty of the nfrastructure. The results are presented n Fgure 4(b). We expected the same relatonshp between the error and the number of sgnals, as n the case of floor detecton. And, ndeed, the accuracy ncreases wth more BTSs that are vsble. However, there s a dfference between the horzontal and vertcal cases, as we can see n Fgure 4(b). The major mprovement n accuracy occurs when we consder sets of fngerprnts wth 2 and 3 BTSs that are vsble. Ths should not be a surprse snce 3 BTSs s the mnmal number needed to provde an exact localsaton n deal, hypothetcal condtons wth no obstacles between the antenna and the recever. On the other hand, there s no bg dfference n accuracy when the number of vsble BTSs ncreases from 4 to 7 (whch s qute dfferent to the case of floor detecton). To sum up, the results presented n Fgures 4(a) and 4(b) show that the greater the number of BTSs that are vsble the hgher the level of accuracy that can be obtaned. Hence, n further consderatons we defne w {,,2,...,7} as the number of BTSs that were reported by the termnal for the fngerprnt F. Toevaluatethenfluenceofthesgnalweghtsusednthe weghted models, we compared three sets of weghts. The frst

7 Moble Informaton Systems Accuracy (%) Error (m) Number of BTSs Number of BTSs Horzontal error Horzontal error for correct floor (a) Floor detecton (b) Coordnates approxmaton Fgure 4: Results wth respect to the number of vsble BTSs Error (%) 6 4 Error (%) Length of path Length of path Equal weghts Sgnal weghts Random weghts Equal weghts Sgnal weghts Random weghts (a) Learnng set (b) Valdaton set Fgure 5: Floor estmaton obtaned by mode flters. set contaned equal weghts. Therefore, the estmatons were calculated accordng to (2) and (4). The second set contaned weghts gven by the number of observed BTSs. Accordng to the results obtaned durng the estmaton for a sngle fngerprnt, the number of observed BTSs s correlated wth the accuracy obtaned. The estmatonswerecalculatedaccordngto(3)and(6). The thrd set contaned random weghts from the doman [,...,7] that covered the range of numbers of observed BTSs. Ths set was a reference set to check that the number of observed BTSs was a better coeffcent than a random value. Estmatons for the floor and coordnates were made separately usng dfferent tendency flters Floor Detecton. In ths secton we perform the analyss of the localsaton soluton descrbed above for the problem of floor predcton. The localsaton s based on k {,...,9} prevous RSS readngs and we consder three varants. Two of them correspond to the floor classfcaton method that followsfrom(2)and(3).thethrdonessmlarto(3)but the weghts w are randomly chosen from the set {,...,7}. The localsaton model was traned usng the S -learnng set and then optmsed usng the S 2 -valdaton set. The accuracy obtaned for the learnng and the valdaton sets s presented n Fgures 5(a) and 5(b), respectvely. In all cases we see that the weghted mode wth weghts correspondng to the vsble BTSs brngs the best results.

8 8 Moble Informaton Systems Part of observatons Part of observatons Error [floor] Error [floor] Length : accuracy: 57.7; Length 2: accuracy: 55.44; Length 3: accuracy: 59.8; Length 4: accuracy: 6.25; Length 5: accuracy: 6.2; Length : accuracy: 57.7; Length 2: accuracy: 55.44; Length 3: accuracy: 59.67; Length 4: accuracy: 6.2; Length 5: accuracy: 62.76; (a) Mode (b) Weghted mode Fgure 6: Accuracy for floor estmaton. For the valdaton seres and the testng seres, the weghted methods provde a substantal mprovement when compared to localsaton based on a sngle fngerprnt (path length k=). Only for the learnng set we obtan worse results as the length of the path consdered for localsaton ncreases. However, ths s somethng one can expect. What s nterestng s that for k>4the random weghts gave better results than the equal weghts. Probably t s the effect of random reducton of the nfluence of the farthest ponts on the path. Fgure 5(b) shows that the results obtaned by the mode flter are best for k [5,6].Wefxk to be 5 as, obvously, the smaller k s the better choce. Afterthelearnngprocessandthevaldaton,theobtaned model was tested on the S 3 -testng set. Fgure 6(a) shows that the mode flter ncreased the accuracy from 57 percent to 6 percent. Fgure 6(b) shows that the weghted mode flterobtanedaccuracyonthelevelof62.8percent.that mproves the accuracy of floor detecton usng weghted mode compared to detecton based on a sngle fngerprnt Coordnates Approxmaton. Smlarly to the prevous secton, we perform the analyss of the localsaton soluton for the problem of horzontal localsaton (the estmaton of x and y coordnates) that reles on k {,...,9} prevous RSS readngs. As before, we consder three varants. Two of them correspond to (4) (7) descrbed n Secton The thrd one s smlar to (6)-(7) but the weghts w are randomly chosen from the set {,...,7}. The medan horzontal error presented n Fgures 7(a) and 7(b) shows that the weghted model gave the best results provded that the number of observed sgnals was taken nto consderaton. The results for random weghts are the worst. As before, for the learnng set we obtan worse results when k ncreases (see Fgure 7(a)). However, Fgure 7(b) shows that the results for the valdaton and the testng set mprove wth the length of the path and hence we obtan an accuracy mprovement when compared to the method basedonasnglefngerprnt(k = ). The best results can be obtaned for k [6,7]. Fgures 8(a) and 8(b) compare results obtaned by algorthms for k {,...,5}. For paths of length 6, the horzontal mean error vares from 6.75 metres for a sngle fngerprnt based method (k =)to5.4metres(k=6). We can also observe more than a 2-metre reducton n the gross errors. The weghted average flter (sgnal weghts w )doesnotbrng bg error reducton n comparson to the average flter (equal weghts). It s about centmetres for both mean and gross errors Aggregaton by a Multlayer Perceptron. The aggregaton process s carred out separately for the floor detecton and the coordnates approxmaton because of the dfferent nput for the neural network. The dfferences le n both the dfferent data and the dfferent structure of the nput layer Floor Detecton. The MLP can be used n several ways nthefloordetectontask.frst,thenetworkcanbeusedto estmate the current floor on the bass of sngle estmatons. Second, the MLP can aggregate the results of the prevous steps. Thrd, both nputs can be merged to create a new model. In our test the MLP brngs 59.54, 64.2, and percent levels of accuracy for the estmaton on the bass of sngle estmatons, the aggregaton, and the merged model, respectvely. Ths shows that the proposed soluton works better than neural network modellng. It also suggests that the results obtaned by weghted model are strongly dfferent

9 Moble Informaton Systems Medan error (m) Medan error (m) Length of path Length of path Equal weghts Sgnal weghts Random weghts Equal weghts Sgnal weghts Random weghts (a) Learnng set (b) Valdaton set Fgure 7: Error for localsaton obtaned by average flters Part of observatons Part of observatons Error (m) Error (m) Length : mean: 6.75; medan: 5.66; 8%: 9.86; Length 2: mean: 5.88; medan: 5.4; 8%: 8.39; Length 3: mean: 5.59; medan: 4.83; 8%: 7.95; Length 4: mean: 5.46; medan: 4.78; 8%: 7.78; Length 5: mean: 5.4; medan: 4.83; 8%: 7.63; Length 6: mean: 5.4; medan: 4.83; 8%: 7.63; Length 7: mean: 5.42; medan: 4.86; 8%: 7.66; Length : mean: 6.75; medan: 5.66; 8%: 9.86; Length 2: mean: 5.8; medan: 4.9; 8%: 8.42; Length 3: mean: 5.5; medan: 4.72; 8%: 7.82; Length 4: mean: 5.32; medan: 4.67; 8%: 7.6; Length 5: mean: 5.32; medan: 4.67; 8%: 7.5; Length 6: mean: 5.3; medan: 4.69; 8%: 7.54; Length 7: mean: 5.32; medan: 4.69; 8%: 7.65; (a) Equal weghts (b) Sgnal weghts Fgure 8: Horzontal error cumulatve dstrbuton. from the results obtaned by the model based on sngle estmatons, and so the merged model cannot work properly. Fgure 9 presents the results obtaned by the aggregatng neural network and Table 4 presents the results for ndvdual floors. The network gves good results for the frst three floors. Fortheupperfloors,resultsareworse,butthemansourceof errors s the thrd floor, where accuracy s about 4 percent. Addtonally, for the 4th and 5th floors, the mean error s over one floor Coordnates Approxmaton. The MLP for the horzontal coordnates approxmaton was tested smlar to the floor detecton task on three dfferent nputs: for separate estmaton, for aggregaton, and for the merged model.

10 Moble Informaton Systems Part of observatons Part of observatons Error [floor] Error (m) Length : accuracy: 56.54; Length 2: accuracy: 53.45; Length 3: accuracy: 58.86; Length 4: accuracy: 59.38; Length 5: accuracy: 63.52; Fgure 9: Accuracy for floor estmaton obtaned by MLP. Table 4: MLP accuracy for ndvdual floors. Floor Accuracy [%] Mean error [floor] Total In our test, the MLP brngs a 6.8, 5.22, and 5.8 mean error for estmaton on the bass of sngle estmatons, aggregaton,andthemergedmodel,respectvely. The errors dmnsh accordng to the ncrease of the length of the paths. However, the length was fxed at 6 n theprevoustests.themeanerror,themedanerror,andthe gross error are less than those for the weghted average flter. Fgure presents the results obtaned by the aggregatng neural network and n Table 5, the results for the ndvdual floors are collected. The obtaned error s not correlated wth the floor error (Table 4) nor the number of ponts. The second floor, whch has the hghest number of measurements, has an average error. The best results are obtaned for the 4th floor, whch has a relatvely hgh floor classfcaton error. The ground floor has the lowest error n the floor classfcaton task. The same floor has one of the worst results n the coordnates estmaton task. The worst results are for the 5th floor. 5. Analyss of Major Accuracy Factors In ths part, we dscuss the major factors that may nfluence the accuracy of the localsaton algorthm. In order to do that, we wll look at the frst step of our algorthm (see Secton 3. pont estmaton) snce the next steps strongly rely on the frst one. We call ths step pont estmaton and t s the Length : mean: 6.6; medan: 5.49; 8%: 9.77; Length 2: mean: 5.88; medan: 5.24; 8%: 8.36; Length 3: mean: 5.39; medan: 4.64; 8%: 7.64; Length 4: mean: 5.43; medan: 4.62; 8%: 7.54; Length 5: mean: 5.32; medan: 4.55; 8%: 7.37; Length 6: mean: 5.8; medan: 4.39; 8%: 7.47; Length 7: mean: 5.7; medan: 4.27; 8%: 7.25; Fgure : Horzontal error cumulatve dstrbuton for MLP. Table 5: MLP pont localsaton errors for ndvdual floor. Floor Mean Medan 8% Total algorthm that provdes localsaton based on a sngle RSS readng (fngerprnt). Let us recap; we tested several statstcal models (see Sectons 4.. and 4..2) and the method based on Baggng performed best. It can be seen that the man problem s the floor predcton whch s stll a subject for further mprovement. On the other hand, the horzontal performance (estmaton of the x and y coordnates) s suffcently good to consder for possble applcatons. The horzontal dstance error cumulatve dstrbuton s gven n Fgure. One can see that the results of the algorthm are slghtly better for the test fngerprnts n the stuaton where the floor coordnate was predcted correctly. Let us also menton that the error for the x coordnate s smaller than the error for the y coordnate. However, ths phenomenon should be expected as the range of x coordnate s sgnfcantly smaller than that for y coordnate. In the remander of ths secton, we shall try to dentfy the factors that may cause the errors descrbed above. 5.. Number of Sgnals. In Sectons 4.. and 4..2, we dscussed the algorthm s performance wth respect to the

11 Moble Informaton Systems Table6:MeanerrorsanalysswthrespecttothenumberofvsbleBTSs. Vsble BTSs Dstrbuton 9% 2% 9% 8% 22% 25% 5% Effectveness (floor) 27% 63% 47% 6% 63% 64% 7% Floor error Error for x [m] Error for y [m] Horzontal error [m] Results under the assumpton that the floor s predcton was correct Error for x [m] Error for y [m] Horzontal error [m] Part of observatons Error (m) General Floor s good Floor s not good Fgure : Horzontal dstance error cumulatve dstrbuton. nfrastructure densty. The analyss gave us some clue on how to proceed n Secton However, for the sake of completeness, we also provde some exact numbers as readng them from graphs can be qute dffcult. These can be found n Table Floor. In ths secton, we analyse errors wth respect tothefloornumber.itsnotanaturalaccuracyfactor, although the followng table shows that the errors vary wth thefloornumber.weshallattempttofndthereasonforths phenomenon. Table 7 shows statstcal data wth respect to the floor number where the measurements were taken. Frstly, let us look at the effectveness of the floor s localsaton as ths s the most mportant ssue n the localsaton problem. Let us observe that there s a sgnfcant dfference n effectvenessforfloornumbersto2,whereouralgorthm works better, and for floor numbers 3 to 5, where the algorthm obtans worse results. It s not surprsng that the best effectveness s obtaned for floor 2, snce ths s where the most fngerprnts were taken when collectng learnng data.however,wehavenosmpleexplanatonastowhythe effectveness for the other floors s so dfferent despte ther beng equally represented n the collecton of fngerprnts gathered for learnng purposes. Let us now look at Fgure 2(b), where we can see the horzontal dstance error for every floor wth and wthout the assumpton that the floor predcton was correct. Once agan, we see that, for the fngerprnts where vertcal predcton was correct, we get a slght mprovement n the horzontal accuracy Area Type. In ths secton, we analyse the effectveness of the floor predcton as well as the mean horzontal dstance error for the dfferent space types n the buldng. We showed that the more regularly shaped spaces covered by the measurement ponts (POIs) where the fngerprnts were taken were mostly covered by good POIs. Here, by regularty we mean that there s no sgnfcant dfference between wdth and length. On the other hand, long corrdors where fngerprnts were only taken along them (and not n the adjacent rooms) were commonly covered by bad POIs. It should be mentoned that the POIs where fngerprnts were taken for learnng purposes (not testng) cover the same regonsasthepoisusedfortestngpurposes.thsleadsusto theconjecturethattheshapeoftheareacoveredbythepois where fngerprnts are taken s mportant. In other words, f the shape s more compact then we obtan better results. Ths would explan why the effectveness s so dfferent on floors 3 to 5. Usng vsualsaton methods, we are able to fnd the areas of the buldng for whch the localsaton algorthm shouldbemproved.suchmapsalsoshowthatthegoodand bad regons are not randomly dstrbuted. They actually gve us a reasonable partton nto easly dentfed and somehow natural sectors of the buldng Other Factors. Besde the accuracy factors dscussed n thesectonsabove,thereareseveralotherfactorsthatnfluence the accuracy. All ths s well covered n []. However, the most mportant factor s the nstablty of the sgnal strength along the tme axs. Ths mght be caused by changes n the weather condtons or n the occupancy of the buldng, or varous other reasons. It obvously vares from day to day

12 2 Moble Informaton Systems Table 7: Mean errors analyss wth respect to the floor number. Floor Dstrbuton 3% 4% % 3% % % Effectveness (floor) 68% 68% 69% 3% 47% 43% Floor error Error for x [m] Error for y [m] Horzontal error [m] Results under the assumpton that the floor s predcton was correct Error for x [m] Error for y [m] Horzontal error [m] Accuracy (%) Error (m) Floor Floor (a) Effcency of the floor s predcton Horzontal error Horzontal error for correct floor (b) Horzontal dstance error Fgure2:Analysswthrespecttothefloornumber. but can also be observed wthn much shorter perods. It should be ponted out that snce the data collecton was very tme consumng, the measurements were taken on several dfferent days. Ths s somethng that defntely nfluences the accuracy of the algorthm. Fgure 3 shows how the sgnals from the key BTSs vary among the three seres for each floor and from floor to floor, wth each seres coverng the same area of the buldng. Ths may suggest that the classcal fngerprnt approach s nsuffcent for the floor recognton problem. Perhaps GSM-based localsaton should be thought of as a supplementary method rather than a stand-alone method when consderng real lfe applcatons n multfloor envronments. On the other hand, the horzontal accuracy of our algorthm seems to be already suffcent for localsaton solutons n large sngle storey buldngs. 6. Comparson wth Other Methods The results obtaned are compared wth the results of the random algorthm as well as results presented n other works. 6.. Comparson wth Random Selecton. Reference levels for the descrbed algorthm were defned by the followng algorthm. A set consstng of (x, y, floor) trplets was created. The set contaned trplets from all measurements from the frst and second seres. Next, paths were cut on overlappng subsequences of length k startng from the kth observaton. Ths testng set was created from paths defned by measurements from the thrd seres; that s, the testng setwasthesameasntheevaluatonofmodeandweghted algorthms descrbed above. At each pont the last pont n the subsequence of the testng path the algorthm produced a randomly selected trplet from the set of learnng trplets defned above. The mean error obtaned for floor dentfcaton was 8 percent. The mean errors for coordnate estmaton were 8.97 and 2.82 metres for x and y, respectvely.thetotalmean error for a pont was 7.2 metres. Ths test proved that the results obtaned by our algorthm are reasonable Comparson wth Other Works. Poston fndng methods and ther accuracy have become mportant ssues n research

13 Moble Informaton Systems (a) Sgnals for floor n seres (d) Sgnals for floor n seres (b) Sgnals for floor n seres (e) Sgnals for floor n seres (c) Sgnals for floor n seres (f) Sgnals for floor n seres (g) Sgnals for floor 2 n seres (j) Sgnals for floor 3 n seres (m) Sgnals for floor 4 n seres (h) Sgnals for floor 2 n seres (k) Sgnals for floor 3 n seres (n) Sgnals for floor 4 n seres 2 Fgure 3: Contnued () Sgnals for floor 2 n seres (l) Sgnals for floor 3 n seres (o) Sgnals for floor 4 n seres 3

14 4 Moble Informaton Systems (p) Sgnals for floor 5 n seres (q) Sgnals for floor 5 n seres 2 Fgure 3: Sgnals strength for BTSs (r) Sgnals for floor 5 n seres 3 over recent years. Outdoor solutons are manly based on GPS/AGPS combned wth a GSM approach. Regrettably, they are not sutable for ndoor envronments because of sgnfcant loss n the GPS sgnal strength. Therefore, many dfferent methods for postonng n ndoor stuatons have been proposed: [9, 7]. These employ dfferent technologes, for example, WF, Bluetooth, nfrared, and varous algorthms for poston calculaton, usng proxmty, dstance or angle estmaton, and complex analyss of the scene. A comprehensve study of the exstng wreless ndoor postonng solutons and ther classfcaton s presented n papers [8, 9]. In addton, [2] comprses an n-depth revew of dfferent methods, tools, and technologes, as well as promsng areas of research for ndoor trackng Localsaton wth GSM Networks. Almost all the exstng ndoor postonng algorthms are based on fngerprntng. In ths secton, we concentrate on methods based on wdely avalable GSM networks. Ths approach does not requre addtonal nstallaton and mantenance, and thus t has been shown to be promsng for locatng moble termnals nsde buldngs. The man lmtng factors here are multpath andfadng.thepowerlevelofthesgnalastsrecevedfrom Access Ponts at a fxed locaton may change due to the user s orentaton or envronmental changes. We focus above all on the accuracy of poston approxmaton, measured n metres, and the requrements (e.g., nput data) for the methods that have been developed. An accurate GSM ndoor localsaton system n multstoreybuldngssproposedn[2].itsbasedonwdesgnal strength fngerprnts and readngs from more than 6 of the strongest cells that are used n the GSM standard. The GSM fngerprnts were collected n a dense grd wth metre to.5 metres of granularty. Four methods, dfferng n the structure of ther fngerprnts (WLAN, the strongest GSM cell, the 6 strongest GSM cells, and up to 35 GSM channels) and the Kmean clusterng algorthm, were mplemented. Experments verfed that ths method s comparable to the WLAN approach (achevng an accuracy of between 2.2 metres and 4.8 metres). The medan accuracy acheved for multfloor postonng was 5 metres; and on just one floor, t reached metres. For the 6 strongest cells, t ranged from 3.4 metres to metres dependng on the envronmental propertes, such as the materal the buldng was made of. Ths proved that extendng fngerprnts, ncludng sgnal strength nformaton from channels other than the 6 strongest cells, can sgnfcantly ncrease localsaton accuracy and, n addton, can dfferentate between floors n both wooden and steelrenforced concrete buldngs, unlke WLANs, whose sgnals are not suffcently weakened by wooden structures between floors. The percentage of erroneous floor classfcaton vared from 2 to 65 percent for less than fngerprnts and decreasedto3 percentforover3gsmchannels. In an ndoor postonng system presented n [22], both GSM and WLAN sgnals were used to estmate the termnal poston usng the nearest neghbour method. Three metrcs Eucldean, Mahalanobs, and probablty (based on the maxmum lkelhood estmator) measurng the dstance betweenthefngerprntsngsm,wlan,andcomposed sgnal data spaces were nvestgated. Tests were performed n an ndoor offce of about 38 metres 42 metres and were lmted to one floor (the 7th of a 9-storey buldng). Both WLAN and GSM data were collected at the same tme. The sgnal strength database was bult wth 8 measurements of the sgnal strength for 6 dstnct ponts along the corrdor wth spacng of about.5 metres and 4 orentatons were collected. The tranng set conssted of 8 percent of randomly chosen data. Expermental results showed that the method usngbothgsmandwlannetworkssstableandcanreach centmetre-level accuracy, whch outperforms the system that utlses a sngle network only. When usng only GSM data, themeanerrorforeucldean,mahalanobs,andprobablty dstances was as low as.69 metres,.87 metres, and.65 metres, and n the worst cases 37.3 metres, 6.76 metres, and 9.4 metres, respectvely. The mean error seems to be over 7 tmes better than that n our method; however, the case tested here was much smpler. The sample ponts formed a polygonal path, smlar to the letter P. Therefore, matchng the estmated ponts s performed n practce n -dmensonal space, not 2-dmensonal or even 3-dmensonal space, as n our algorthm. The effectveness of GSM-based localsaton methods depends on the number of avalable and examned GSM carrers. The work [23] presents an algorthm based on fullband GSM fngerprnts. Tests were performed n an urban 5- room apartment on the 5th (top) floor. Both the RSS and base staton dentty code (BSIC) were recorded twce a day for a month, for the full avalable set of 498 GSM carrers usng the

15 Moble Informaton Systems 5 TEMS trace moble system. The data set contaned 24 scans assgned to 5 rooms. The tranng set ncluded 69 examples, and the valdaton set ncluded 72 examples. Three types of classfers were compared: nearest neghbour, support vector machne (SVM), and Gaussan process. To reduce the complexty of classfers, the 3 fngerprnt types (wth a lmted number of carrers wth the strongest RSS values) were defned and were then compared to the one contanng all GSM carrers. The commonly used 7-carrer-based measurements classfed percent of fngerprnts correctly, and the effectveness for fngerprnts of length 35 was percent. Very good performance (97.8 percent of correct rado fngerprnt classfcatons) was obtaned n the case of the lnear SVM method for all actve carrers, and RSS values were ncluded. It should be noted that the localsaton s lmted to one floor and to room-level only; but the results confrm that the greater the extent of the fngerprnts avalable, the hgher the accuracy provded. Further extensons of ths method combned wth nertal sensors and a ste map are the subject of [24], where an overall 7 percent correct classfcaton was obtaned wth msclassfcatons coverng manly the adjacent rooms. The authors report that localsaton was performed n a so-called short tme wth a very small tranng set. The problem of choosng a subset of relevant GSM carrers provdng a good dstncton between rooms s also examned n [25]. Measurements were also carred out over one month n 5 of 8 rooms of a 2nd-floor offce, usng a TELIT GM-862 modem that detected 534 dfferent carrers. The nput data conssted of 6 measurements assgned to 5 rooms. Two algorthms for rankng nput varables were nvestgated: forward regresson usng Gram-Schmdt orthogonalzaton and SVM recursve feature elmnaton. The results obtaned proved that the 6 most relevant carrers aresuffcenttoproperlylocalse97percentofscansnan ndependent test set by both algorthms Floor Classfcaton. Floor classfcaton tself s the subjectofresearch[26].thepaperpresentsskyloc,agsm fngerprntng-based system, whch determnes the current floor on whch a user wth a moble phone s located. The algorthm that predcts the floor uses the smallest Eucldean dstance. To mprove performance and memory usage, the sze of fngerprnts has been reduced; napproprate sgnal sourceshavebeenelmnatedvathefollowngfeatureselecton technques: forward selecton, backward elmnaton, and a new per-floor selecton. The next optmsaton was acheved usng the sldng wndow algorthm, whch frst classfes each measurement ndvdually, and then pcks the current floor as the most frequently appearng floor among a sequence of 5,, or 2 results. The tests were conducted n 3 tall (9-, 2-, and 6-storey) buldngs, usng as the data set, 3, and 3 fngerprnts per floor, respectvely. The fngerprnts were collected along user paths wth a spacng of about 2 metres wth dfferent hardware separately for tranng and testng sets, at an nterval of 2 days or a month, dependng on the buldng. For 2 testng ponts n a sldng wndow, the system classfed the floor correctly n 5 percent to 73 percent of cases. For 5 testng ponts, compared to our tests of 4 ponts of hstorcal estmatons and 64 percent accuracy, the obtaned results were worse the system correctly classfed from 45 percent to 6 percent of cases. Surprsngly, the number of tranng fngerprnts per floor had lttle mpact on the localsaton accuracy. As we verfed before, the effectveness of 3-dmensonal localsaton may be mproved when the phases of recognton of the floor, and the estmaton of 2-dmensonal coordnates wthn a gven (sngle) floor, are separated. Floor detecton can be sgnfcantly enhanced by usng addtonal sensors embedded n moble devces. The work [27] descrbes a combned seamless 3-dmensonal localsaton system usng an atmospherc pressure sensor to calculate the user s alttude, andthusthefloor,nrealtme.thetestswerecarredoutna 4-storey buldng. The pressure accuracy was better than 2 Pa, gvng an alttude resoluton of better than.5 m; hence, the system could accurately determne the floor. The paper [28] presents a method of usng a moble phone s accelerometer only and analysng user trals, to map the current floor level. It does not requre any nfrastructure or any pror nformaton of the buldng. A feld study nvolved moble users for three hours n a -floor buldng wth two elevators. Each smartphone was equpped wth an embedded 3-axs accelerometer. The accuracy approached 9 percent n 2 hours and reached 97 percent n three hours. The authors reported further mprovements n the localsaton based on crowdsourcng n the work [29]. Then, [3] descrbes system leveragng also usng crowdsourcng and buldng a barometer fngerprnt map contanng the atmospherc pressure value for each floor level. The expermental results showed a very hgh accuracy for ths method, of over 98 percent. In all the works descrbed, the complexty of the proposed solutons s not gven. The authors have wrtten that algorthms work n short tme or even n real tme. Only n [26] was performance evaluaton performed. On average, t took.2 seconds to match a sngle testng fngerprnt to a sngle tranng fngerprnt; therefore, matchng a current moble phone fngerprnt to a whole tranng set would take about a second. To sum up, the errors for our soluton (4.62 metres for a horzontal localsaton n a multstorey buldng and 64 percent for the floor detecton) are comparable to, or even better than, those obtaned n the revewed works and based only on receved GSM sgnal strength. The results reported n lterature for methods usng addtonal sensors n a moble system seem to be a step n the rght drecton towards mprovng the accuracy of our localsaton system. 7. Conclusons Ths work descrbes and analyses a localsaton soluton n a sx-floor academc buldng. The localsaton was based on the sgnalsfromglobalsystemformoblecommuncatonsbase Transcever Statons. The localsed objects were common moble phones. We demonstrated that, for the localsaton based on a sngle fngerprnt, we obtaned a horzontal medan error of about 6.5 metres and accuracy of floor detecton of about 56%. We proposed a three-step method that uses not only a sngle fngerprnt but also the precedng

16 6 Moble Informaton Systems Table 8: Summary of the localsaton results of ndvdual steps of the localsaton process. Stage Horzontal error Floor Mean [m] Medan [m] 8% [m] Error [%] Estmaton for sngle fngerprnt Estmaton for fngerprnts sequence MLP aggregaton ones. Ths method mproved the accuracy. We obtaned a horzontal medan error of around 4.4 metres and floor detecton accuracy of over 64%. Table 8 summarses the errors obtaned durng the localsaton process for each step separately. It shows that ths method mproves localsaton n every case. The localsaton error for coordnates obtaned n our workssmlartotheresultsobtanedndfferentworks that used addtonal nformaton such as WLANs sgnals or accelerometers data. A medan error of below 5 metres s acceptable n ndoor localsaton. However, the accuracy of the floor detecton algorthm s relatvely low. Ths problem can be solved usng WLANs sgnals. Our tests proved that, by usng a combnaton of GSM and WLAN sgnals, the current floor can be detected wth an accuracy of over 9 percent. Ths work focuses on GSM sgnals and n ths case a pure GSM sgnal seems to be nsuffcent to localse a floor n a multstorey buldng. In further works we want to collect addtonal data from the same and other buldngs to examne long-term dfferences n the sgnal map, collect addtonal data such as accelerometer data, and apply the proposed methods for varous buldngs. Competng Interests The authors declare that they have no competng nterests. Acknowledgments The authors would lke to thank Le Dnh Tung, Łukasz Wrzesńsk, and Bogusław Zaręba for makng the measurements and Orange Labs Poland for provdng the cell phones andtheapplcatonfortheandrodos.theresearchs supported by the Natonal Centre for Research and Development, Grant no. PBS2/B3/24/24, applcaton no References [] 3GPP, Moble rado nterface layer 3 specfcaton; rado resource control protocol, 3GPP TS 4.8 V8.27. (26-5), 3GPP, 26. [2] M. Grzenda, On the predcton of floor dentfcaton credblty n rss-based postonng technques, n Recent Trends n Appled Artfcal Intellgence, M.Al,T.Bosse,K.V.Hndrks, M. Hoogendoorn, C. M. Jonker, and J. Treur, Eds., vol. 796 of Lecture Notes n Computer Scence, pp. 6 69, 23. [3] J. Karwowsk, M. Okulewcz, and J. Legersk, Applcaton of partcle swarm optmzaton algorthm to neural network tranng process n the localzaton of the moble termnal, n Engneerng Applcatons of Neural Networks, vol.383 of Communcatons n Computer and Informaton Scence,pp.22 3, Sprnger, Berln, Germany, 23. [4] D.Dellng,P.Sanders,D.Schultes,andD.Wagner, Engneerng route plannng algorthms, n Algorthmcs of Large and Complex Networks, J. Lerner, D. Wagner, and K. A. Zweg, Eds., vol. 555 of Lecture Notes n Computer Scence, pp. 7 39, Sprnger, Berln, Germany, 29. [5] Y. Freund and R. E. Schapre, A decson-theoretc generalzaton of on-lne learnng and an applcaton to boostng, Journal of Computer and System Scences,vol.55,no.,pp.9 39,997. [6] L. Breman, Random forests, Machne Learnng, vol. 45, no., pp. 5 32, 2. [7] T. Haste, R. Tbshran, and J. Fredman, The Elements of Statstcal Learnng, Sprnger Seres n Statstcs, Sprnger, New York, NY, USA, 2. [8] R. E. Schapre and Y. Snger, Improved boostng algorthms usng confdence-rated predctons, n Machne Learnng, pp. 8 9, 999. [9] F. Evennou and F. Marx, Advanced ntegraton of WF and nertal navgaton systems for ndoor moble postonng, EURASIP Journal on Appled Sgnal Processng,vol.26,Artcle ID 8676, 26. [] C.Bento,T.Soares,M.Veloso,andB.Baptsta, Astudyonthe sutablty of GSM sgnatures for ndoor locaton, n ProceedngsoftheEuropeanConferenceonAmbentIntellgence(AmI 7),pp.8 23,Sprnger,27. [] F. Zampella, A. R. Jmenez Ruz, and F. Seco Granja, Indoor postonng usng effcent map matchng, rss measurements, andanmprovedmotonmodel, IEEE Transactons on Vehcular Technology,vol.64,no.4,pp.34 37,25. [2]D.Pastna,F.Colone,T.Martell,andP.Falcone, Parastc explotaton of W-F sgnals for ndoor radar survellance, IEEE Transactons on Vehcular Technology, vol.64,no.4,pp. 4 45, 25. [3] P. Wawrzynak, P. Korbel, P. Skulmowsk, and P. Poryzala, Mxed-mode wreless ndoor postonng system usng proxmty detecton and database correlaton, n Proceedngs of the Federated Conference on Computer Scence and Informaton Systems (FedCSIS 4),pp.35 42,September24. [4] C. Wu, Z. Yang, and Y. Lu, Smartphones based crowdsourcng for ndoor localzaton, IEEE Transactons on Moble Computng,vol.4,no.2,pp ,25. [5]K.Pwowarczyk,P.Korbel,andT.Kacprzak, Analyssofthe nfluence of rado beacon placement on the accuracy of ndoor postonng system, n Proceedngs of the Federated Conference on Computer Scence and Informaton Systems (FedCSIS 3),pp , Kraków, Poland, September 23. [6] A. Papapostolou and H. Chaouch, Scene analyss ndoor postonng enhancements, Annals of Telecommuncatons,vol. 66,no.9-,pp ,2.

17 Moble Informaton Systems 7 [7] Y.X.Zhao,Q.Shen,andL.M.Zhang, Anovelhghaccuracy ndoor postonng system based on wreless lans, Progress n Electromagnetcs Research C,vol.24,pp.25 42,2. [8] H. Lu, H. Darab, P. Banerjee, and J. Lu, Survey of wreless ndoor postonng technques and systems, IEEE Transactons on Systems, Man and Cybernetcs Part C: Applcatons and Revews, vol. 37, no. 6, pp. 67 8, 27. [9] Y. Gu, A. Lo, and I. Nemegeers, A survey of ndoor postonng systems for wreless personal networks, IEEE Communcatons Surveys and Tutorals, vol., no., pp. 3 32, 29. [2] D. Dardar, P. Closas, and P. M. Djurc, Indoor trackng: theory, methods, and technologes, IEEE Transactons on Vehcular Technology,vol.64,no.4,pp ,25. [2] V.Otsason,A.Varshavsky,A.LaMarca,andE.deLara, Accurate GSM ndoor localzaton, n UbComp 25: Ubqutous Computng: 7th Internatonal Conference, UbComp 25, Tokyo, Japan, September 4, 25. Proceedngs, vol.366oflecture Notes n Computer Scence, pp. 4 58, Sprnger, Berln, Germany, 25. [22] J.Zhou,W.M.-C.Yeung,andJ.K.-Y.Ng, Enhancngndoor postonng accuracy by utlzng sgnals from both the moble phone network and the wreless local area network, n Proceedngs of the 22nd Internatonal Conference on Advanced Informaton Networkng and Applcatons (AINA 8), pp.38 45, Oknawa, Japan, March 28. [23] B. Denby, Y. Oussar, I. Ahrz, and G. Dreyfus, Hgh-performance ndoor localzaton wth full-band GSM fngerprnts, n Proceedngs of the IEEE Internatonal Conference on Communcatons Workshops (ICC 9), pp. 5, Dresden, Germany, June 29. [24] Y.Tan,B.Denby,I.Ahrz,P.Roussel,andG.Dreyfus, Hybrd ndoor localzaton usng GSM fngerprnts, embedded sensors and a partcle flter, n Proceedngs of the th Internatonal Symposum on Wreless Communcatons Systems (ISWCS 4), pp , Barcelona, Span, August 24. [25] I. Ahrz, Y. Oussar, B. Denby, and G. Dreyfus, Carrer relevance study for ndoor localzaton usng GSM, n Proceedngs of the 7th Workshop on Postonng, Navgaton and Communcaton (WPNC ), pp , Dresden, Germany, March 2. [26] A.Varshavsky,A.LaMarca,J.Hghtower,andE.deLara, The skyloc floor localzaton system, n Proceedngs of the 5th Annual IEEE Internatonal Conference on Pervasve Computng and Communcatons (PerCom 7), pp.25 34,WhtePlans, NY, USA, March 27. [27] N.He,J.Huo,Y.Dong,Y.L,Y.Yu,andY.Ren, Atmospherc pressure-aware seamless 3-D localzaton and navgaton for moble nternet devces, Tsnghua Scence and Technology, vol. 7,no.2,pp.72 78,22. [28] H. Ye, T. Gu, X. Zhu et al., FTrack: nfrastructure-free floor localzaton va moble phone sensng, n Proceedngs of the th IEEE Internatonal Conference on Pervasve Computng and Communcatons (PerCom 2), pp. 2, Lugano, Swtzerland, March 22. [29] H. Ye, T. Gu, X. Tao, and J. Lu, F-Loc: floor localzaton va crowdsourcng, n Proceedngs of the 2th IEEE Internatonal Conference on Parallel and Dstrbuted Systems (ICPADS 4),pp , Hsnchu, Tawan, December 24. [3] H.Ye,T.Gu,X.Tao,andJ.Lu, B-loc:scalablefloorlocalzaton usng barometer on smartphone, n Proceedngs of the IEEE th Internatonal Conference on Moble Ad Hoc and Sensor Systems (MASS 4), pp , Phladelpha, Pa, USA, October 24.

18 Journal of Advances n Industral Engneerng Multmeda The Scentfc World Journal Volume 24 Volume 24 Appled Computatonal Intellgence and Soft Computng Internatonal Journal of Dstrbuted Sensor Networks Volume 24 Volume 24 Volume 24 Advances n Fuzzy Systems Modellng & Smulaton n Engneerng Volume 24 Volume 24 Submt your manuscrpts at Journal of Computer Networks and Communcatons Advances n Artfcal Intellgence Volume 24 Internatonal Journal of Bomedcal Imagng Volume 24 Advances n Artfcal Neural Systems Internatonal Journal of Computer Engneerng Computer Games Technology Advances n Volume 24 Advances n Software Engneerng Volume 24 Volume 24 Volume 24 Volume 24 Internatonal Journal of Reconfgurable Computng Robotcs Computatonal Intellgence and Neuroscence Advances n Human-Computer Interacton Journal of Volume 24 Volume 24 Journal of Electrcal and Computer Engneerng Volume 24 Volume 24 Volume 24