Pervasive and Mobile Computing. Adaptive radio maps for pattern-matching localization via inter-beacon co-calibration

Transcription

1 Pervasve and Moble Computng ( ) Contents lsts avalable at ScVerse ScenceDrect Pervasve and Moble Computng journal homepage: Adaptve rado maps for pattern-matchng localzaton va nter-beacon co-calbraton Ch-Chung Lo, Lan-Yn Hsu, Yu-Chee Tseng Department of Computer Scence, Natonal Chao-Tung Unversty, Hsn-Chu, 30010, Tawan a r t c l e n f o a b s t r a c t Artcle hstory: Receved 24 February 2011 Receved n revsed form 30 December 2011 Accepted 2 January 2012 Avalable onlne xxxx Keywords: Localzaton Locaton-based servce Trajectory trackng Wreless sensor network A growng number of locaton-based applcatons are based on ndoor postonng, and much of the research effort n ths feld has focused on the pattern-matchng approach. Ths approach reles on comparng a pre-traned database (or rado map) wth the receved sgnal strength (RSS) of a moble devce. However, such methods are hghly senstve to envronmental dynamcs. A number of solutons based on added anchor ponts have been proposed to overcome ths problem. Ths paper proposes an approach usng exstng beacons to measure the RSS from other beacons as a reference, whch we call nter-beacon measurement, for the calbraton of rado maps on the fly. Ths approach s feasble because most current beacons (such as W-F and ZgBee statons) have both transmttng and recevng capabltes. Ths approach would releve the need for addtonal anchor ponts that deal wth envronmental dynamcs. Smulaton and expermental results are presented to verfy our clams Elsever B.V. All rghts reserved. 1. Introducton Locaton-based servces (LBS) are hghly popular applcatons n moble and wreless technology [1], and locaton trackng s the core problem [2 4]. GPS s currently the most wdely used technology for locaton trackng n outdoor envronments; however, due to effects such as shadowng, GPS cannot be used ndoors. Recently, consderable research has been dedcated to the development of wreless networks as an nfrastructure for ndoor locaton trackng. One promsng approach s the pattern-matchng technque [5 8], whch s capable of meter-level accuracy. It uses the receved sgnal strength (RSS) of the rado frequency emtted by the statons of wreless networks as reference. We call these statons beacons. A pattern-matchng system usually works n two phases: tranng and postonng. In the tranng phase, the operator collects the RSS of beacons at varous tranng locatons to form a database referred to as a the rado map. Durng the postonng phase, the moble devce compares ts current RSS aganst the rado map to determne ts locaton. Fluctuatons resultng from envronmental dynamcs are the major drawback of the pattern-matchng approach. Temperature, humdty, and movng objects are all capable of dsruptng the observed RSS, leadng to devatons n the traned rado map. LANDMARC [9] s a soluton relyng on actve RFIDs and tranng data obtaned from onlne sources to perform ndoor localzaton. LEMT [10] s another approach usng anchor ponts to derve adaptve temporal rado maps to overcome ths problem. However, all of these exstng solutons rely on addtonal hardware to deal wth envronmental dynamcs and to calbrate the rado maps. Ths paper proposes a novel approach that allows for the self-calbraton of rado maps wthout the need for addtonal hardware. Most current beacons (such as W-F APs and ZgBee nodes) have both transmttng and recevng capabltes. Correspondng author. E-mal addresses: ccluo@cs.nctu.edu.tw (C.-C. Lo), lyhsu@cs.nctu.edu.tw (L.-Y. Hsu), yctseng@cs.nctu.edu.tw (Y.-C. Tseng) /$ see front matter 2012 Elsever B.V. All rghts reserved. do: /j.pmcj

2 2 C.-C. Lo et al. / Pervasve and Moble Computng ( ) Usng these beacons to observe the RSS from other beacons, whch we call nter-beacon measurement, would enable the capture of current envronmental dynamcs and the calbraton of rado maps on the fly. The remander of ths paper s organzed as follows. In Secton 2, we revew some related works. In Secton 3, we propose a framework to enable nter-beacon measurement as well as two schemes to calbrate rado maps, based on data clusterng and data regresson. In Secton 4, we verfy our results va smulatons and practcal experments. Conclusons are drawn n Secton Background and related works The mplementaton of rado-based wreless networks for ndoor localzaton reles on ether a rado-propagaton model [11] or an emprcal-ft model [12]. In a rado-propagaton model, a mult-lateraton mechansm s requred to calculate the locaton of a devce. The path loss of beacon b at a dstance d s normally modeled as follows [13]: P r (d, b) = P t (b) PL (d 0, b) 10η log 10 d d 0 + N(0, σ ), (1) where P t (b) s the transmttng power of b, PL(d 0, b) s the path loss at dstance d 0, η s an envronment- and hardwaredependent constant, d 0 s a reference dstance, and N(0, σ ) s a zero-mean normal-dstrbuton random varable. Unfortunately, ths model s not adequate for ndoor envronments wth dynamcally changng η and σ. Ths study focuses on the emprcal-ft model [5,14,15], also known as pattern-matchng localzaton. Gven a set of beacons B = {b 1, b 2,..., b m } and a set of tranng locatons L = {l 1, l 2,..., l n } n a sensng feld, ths method proceeds through two phases. In the tranng phase, we measure the RSS vectors of all beacons at each tranng locaton l over a perod of tme and create a feature vector υ = [υ,1, υ,2,..., υ,m ] for l, where υ,j R s the average RSS from b j, j = 1... m. These feature vectors are collected n a set V = {υ 1, υ 2,..., υ n }, called the rado map. In the postonng phase, a moble devce measures ts current RSS vector s = [s 1, s 2,..., s m ] and compares s aganst V. The best matched one or ones n V are used to predct the current locaton of the devce. In practce, we could also choose the nearest neghbor [5] and use probablstc technques [15] as a matchng method. In [5], the nearest neghbor algorthm s appled to search for the best match accordng to Eucldean dstances n the sgnal space. In [6], a probablstc framework for localzaton s presented to handle fluctuatons n sgnal strength. Generally, the probablstc approach can more accurately reflect the dynamcs assocated wth changes n RSS. Below, we revew exstng works from three perspectves: (1) technques to mprove scalablty, (2) technques to explot hstory trackng nformaton, and (3) technques to handle RSS dynamcs. To mprove scalablty, t should be noted that all pattern-matchng solutons rely on a large amount of tranng data. To deal wth ths ssue, Refs. [16,17] propose clusterng-based methods to reduce comparson costs. The man dea s to apply clusterng technques to dvde a rado map nto smaller sub-maps. To explot hstory trackng nformaton, many researchers have used a sequence of postonng results to mprove accuracy. In [18], Bayesan flters were developed to ntegrate multple sources of sensng data. The key features of these flters are observaton, predcton, and hstory models to remove unrelable data. For example, a trackng system s capable of explotng the moblty hstory of users to speculate as to ther trajectory [19]. By contrast, partcle flters are used n [18] to reflect the probablty denstes of our belefs based on prevous measurements. In [20], the belef about a dynamc system at tme t s represented as a probablty dstrbuton over the state space. To deal wth envronmental dynamcs, [9] used RFID tags as references for RSS dstrbuton to ad n localzaton. In [21], statonary emtters and snffers were used to determne the tranng data and assst n ndoor localzaton onlne. A sensorasssted scheme was proposed n [22] to measure the current rado map. LEMT [10] s the frst research to use anchor ponts to derve adaptve temporal rado maps capable of overcomng envronmental dynamcs. However, these approaches all requred addtonal hardware to deal wth envronmental dynamcs and to calbrate the rado maps. 3. Adaptve rado maps va nter-beacon measurement An nherent lmtaton of the pattern-matchng localzaton method s the problem of sgnal nstablty. To deal wth ths ssue, we propose a method based on nter-beacon measurement. We have observed that most beacons n use have both transmttng and recevng capabltes. Recrutng these beacons to measure the RSS of neghborng beacons, would enable the adaptve calbraton of rado maps. Consder the example n Fg. 1. Suppose that b and b j are two beacons and l s a tranng locaton. In the tranng phase, let S and S j be the RSSs of b and b j measured at l, respectvely. Durng the postonng phase, suppose that a devce at l measures the RSSs of b and b j as S and S j, respectvely. If {S, S j } were to devate too far from {S, S j }, t would be exceedngly dffcult to determne whether the devce s at poston l. The concept behnd the proposed nter-beacon measurement method s to add two tags, S,j and S j,, durng the tranng phase to represent the RSS of b as observed by b j and the RSS of b j as observed by b. It the postonng phase, n addton to measurng S and S j, we also collect S,j (the RSS of b as observed by b j ) and S j, (the RSS of b j as observed by b ). It s expected that usng the set {S, S j, S,j, S j, } collected n the tranng phase and the set {S, S, j S,,j S j, } collected n the postonng phase, would provde addtonal clues to determne whether the moble devce s located near l.

3 C.-C. Lo et al. / Pervasve and Moble Computng ( ) 3 Fg. 1. An example of nter-beacon measurement. Fg. 2. Flow chart of the proposed solutons. In the followng, we formally defne the problem as follows. We defne a set of beacons B = {b 1, b 2,..., b m } and a set of tranng locatons L = {l 1, l 2,..., l n } n a sensng feld. In the tranng phase, suppose that we measure the RSS vectors at each locaton l to create the rado map V n conjuncton wth the measurements among the beacons. The queston s: In the postonng phase, how do we provde an adaptve rado map V, based on the current observatons of nter-beacon measurement to facltate the task of localzaton? We propose two solutons as follows. The frst soluton uses nformaton related to the nter-beacon measurement to cluster tranng data nto multple rado maps. To poston a devce, we frst use the current nter-beacon measurement to select an approprate rado map, from whch we choose the closest locaton. Ths soluton s based on the assumpton that there should be a hgh correlaton between each nter-beacon measurement and ts correspondng tranng cluster. Therefore, the current nter-beacon measurement s a good ndcator of the cluster to be used durng the postonng phase. Conversely, the second soluton nvolves the use of nter-beacon measurements to nterpolate the current rado map. Ths soluton s based on the assumpton that the correlaton between each nter-beacon measurement and ts correspondng tranng data can be predcted usng a lnear regresson model. Both solutons comprse three phases, as llustrated n Fg Soluton 1: Clusterng-based scheme Beacon-asssted tranng phase: In each tranng locaton, we collect two types of RSS: beacon-to-devce RSS (BD-RSS) and beacon-to-beacon RSS (BB-RSS). BB-RSSs reflect the envronmental characterstcs when the correspondng BD-RSSs are collected. Specfcally, multple (BD-RSS, BB-RSS) pars wll be collected at each l. Each BD-RSS s a vector wth the format υ (x) = [υ (x),j ] j=1...m, where x s the tmestamp of when the vector was measured and υ (x),j s the RSS of beacon b j measured at l. When υ (x) s recorded, the system also records the RSS of b j measured by b k, denoted by µ (x),j,k. These measurements

4 4 C.-C. Lo et al. / Pervasve and Moble Computng ( ) are recorded n a BB-RSS vector µ (x) = [µ (x),j,k ] j=1...m,k=1...m,j k. In practce, t s dffcult to ensure that a BD-RSS and a BB-RSS are taken at precsely the same tme, consderng that beacons have regular jobs to perform. A degree of tmng dfference s acceptable as long as the envronment remans roughly smlar between measurements. For smplcty, we stll use the same superscrpt (x) here. The combnaton (υ (x), µ (x) ) s called a (BD-RSS, BB-RSS) par measured at tme x for l. The collectons are mantaned n a tranng database T = {(υ (x), µ (x) ) l L, x}. To enable nter-beacon measurement, each beacon must swtch to receve mode from tme to tme. Ths can be performed easly by modern W-F and ZgBee nterfaces. In addton, to ncrease the dversty of database T, the measurng tme xs should be as dversfed as much as possble. For example, we could conduct measurements on sunny and rany days, on workng days and holdays. Data clusterng phase: Because database T s collected wth dversty n mnd, we suggest parttonng T nto several subsets, each called a rado map, accordng to ther smlarty. Below, we propose a modfed k-means clusterng algorthm [7, 23] to acheve ths goal. 1. Apply the k-means algorthm to partton T nto k subsets usng the BB-RSS of each (BD-RSS, BB-RSS) par as the key. The k-means process nvolves multple data-clusterng teratons n whch these keys are compared. Specfcally, when comparng the smlarty between two (BD-RSS, BB-RSS) pars P = (υ (x) p, µ (x) p ) and Q = (υ (y) q, µ (y) q ), we defne ther dstance as d(p, Q ) = j, k 2. µ (x) p,j,k q,j,k µ(y) Here, a larger dstance means a lower degree of smlarty. Each teraton of the process generates k subsets. Intutvely, pars wth smlar BB-RSSs (.e., those measured under smlar condtons) are nserted nto the same subset. 2. Let the k subsets obtaned n step 1 be T 1, T 2,..., T k. For each T, = 1... k, we defne a feature vector for T based on the BB-RSSs of the members of T. Specfcally, T s feature vector s ω = [ω j,k ] j=1...m,k=1...m,j k, where ω j,k = (υ (x) p,µ (x) p ) T µ (x) p,j,k T. 3. The above defned T, = 1... k, s not necessarly well-formed because a number of the tranng locatons may not appear n T (the k-means algorthm does not guarantee ths property). To ensure that T s a well-formed rado map, we must check whether there exsts at least one (υ (x) p, µ (x) p ) T for each l p L. If not, we compare the ω of T aganst all (BD-RSS, BB-RSS) pars for l p n T. The par for whch BB-RSS s most smlar to ω s added to T. Wth ths amendment, T becomes well-formed. In practce, for a gven total number of t tranng samples, we set k = t. In addton, t should be noted that t s possble that multple unts of tranng data sampled at dfferent tmes n the same tranng locaton may be nserted nto the same rado map. Ths s because they may have smlar envronmental characterstcs. These data may be averaged nto a sngle unt or reman unchanged, dependng on the localzaton algorthm used n the postonng phase. Clusterng-based postonng phase: When a devce s requred to determne ts locaton, t measures ts current BD-RSS vector, denoted by υ c = [ υ c,j ] j=1...m. It then submts υ c to the locaton server, whch takes the followng actons. 1. The locaton server frst requests that all beacons measure the RSS of the others. The collected current BB-RSS vector s denoted by µ = [ µ j,k ] j=1...m,k=1...m,j k. 2. The locaton server then compares µ aganst the ω of each T. The dstance between µ and ω s defned as 2 d( µ, ω ) = µj,k ω j,k. (4) j, k Let T be the one for whch d( µ, ω ) s the smallest. We then select T as the current rado map and compare υ c aganst the BD-RSS of each (BD-RSS, BB-RSS) par n T. When comparng the smlarty of υ c and a BD-RSS υ p, we defne ther dstance as 2 d( υ c, υ p ) = υc,j υ p,j. (5) j The locaton for whch the correspondng BD-RSS s most smlar to υ c s estmated as the locaton of the devce. In Step 2, a BB-RSS must be measured n response to every locaton query. To reduce overheads, the locaton server can perodcally collect BB-RSS vectors, such that the most recent s regarded as µ. (2) (3)

5 C.-C. Lo et al. / Pervasve and Moble Computng ( ) Walls Path1 Path2 Path3 30 Y axs X axs Fg. 3. The smulated envronment Soluton 2: Regresson-based scheme The Beacon-asssted tranng phase s the same; therefore, we wll only dscuss the followng two phases. Data regresson phase: Recall that for each tranng locaton l p, we have already collected a number of (BD-RSS, BB-RSS) pars n T. Gven the current BB-RSS, we can use these pars to predct the current BD-RSS vector at l p usng a regresson method. Let T p be the set of (BD-RSS, BB-RSS) pars collected at l p. We assume the followng lnear relaton: υ (x) p, = x, j,j a j µ (x) p,,j + b. Intutvely, we use terms on the rght-hand sde to predct the RSS of b measured at l p. Eq. (6) can be establshed for each par n T p, resultng n µ (1) p,,1 µ (1) p,,m 1.. µ (x) p,,1 µ (x) p,,m 1 A p, a 1... a m b Usng least-squares analyss, we obtan B p, = (A T p, A p,) 1 A T p, C p,. B p, = υ (1) p,. υ (x) p, C p,. (7) Note that f we contnue ncreasng the number of tranng vectors, B p, wll also be changed. In practce, we can update B p, perodcally to balance the computng cost and postonng accuracy. In addton, the sze of T p should be bounded to ensure that computng Eq. (8) remans feasble. Regresson-based postonng phase: When a devce must determne ts locaton, t measures ts current BD-RSS vector υ c and submts υ c to the locaton server, whch then takes the followng actons. 1. It frst request all beacons measure the RSS of the others. Let the RSS of b measured by b j be µ,j. 2. Usng the prevously obtaned B p,, the server then predcts the current RSS vector at locaton l p as υ p = [ υ p,1, υ p,2,..., υ p,m ], where υ p, = [ µ,1, µ,2,..., µ,m, 1] B p,, = 1... m. 3. We then compare υ c aganst each υ p for all tranng locatons. The l p whch provdes the smallest dstance d( υ c, υ p ) s estmated as the locaton of the devces. 4. Smulaton and expermental results (6) (8) 4.1. Smulaton results To verfy our results, we smulated a 50 m 50 m sensng feld wth 12 beacons. Each beacon had a rado power of 15 dbm to ensure that the beacons could reach each other. Tranng locatons were grd ponts separated by 1 m, and fve samples were taken from each tranng locaton. To complcate the envronment, a number of vertcal and horzontal walls were placed on the feld, as shown n Fg. 3. Several roamng paths of users were smulated. Note that users may occasonally

6 6 C.-C. Lo et al. / Pervasve and Moble Computng ( ) Fg. 4. Impact of k n the clusterng-based scheme. pass walls; the purpose s to see how the scheme performs when there are sudden sgnal changes. We adopted the path loss model of RIM [24] and rewrote Eq. (1) as follows: P r (d, b) = P t (b) PL DOI (d, b) PL WAF (d, b) + N(0, σ ). (9) In RIM, DOI stands for degree of rregularty, and s used to control the amount of path loss n dfferent drectons, PL DOI (d, b) = PL (d 0, b) + 10η log 10 K, (10) dd0 where K s used to model the level of rregularty at degree ( = ), K = 1 f = 0 K 1 ± W(0, β, φ) DOI f = (11) where K 0 K 359 DOI and W(0, β, φ) s a zero-mean Webull random varable wth a slope parameter β and a scale parameter φ. Here, we let β = 1 and φ = 0.1. The resultng PL DOI (d, b) has non-sotropc and contnuous propertes. To model the mpact of ndoor parttons and walls n such an envronment, we employed a wall attenuaton factor (WAF) [5], PL WAF (d, b) = mn (N obs, N max ) WAF, where N obs s the number of walls crossed by a lne-of-sght path, N max s the maxmum number of walls that can nfluence PL WAF (d, b), and WAF s the amount of sgnal attenuaton caused by a sngle wall. Note that P r (d, b) s a random varable, whch may change at any moment. The total smulaton tme was 1000 s. The movng speed of the users was set to 1 m/s and RSSs were measured every second. The default smulaton parameters were P t (b) = 15 dbm, d 0 = 1 m, σ = 2 or 4, PL(d 0, b) = 37.3 dbm, η = 3.3, DOI = 0.01, WAF = 3, and N max = 4. We compared our scheme aganst the NNSS (nearest neghbor n sgnal space) scheme [5]. We also smulated an deal NNSS scheme, whch assumes that the η used n the tranng phase s known n the postonng phase. For each measurement (ncludng tranng and postonng phases), we randomly selected η n [3.1, 3.5] and [2, 4] to reflect the envronmental dynamcs. Below, we dscuss our smulaton results from four perspectves. (1) The mpact of k n the clusterng-based scheme: Fg. 4 shows the mpact of k on postonng accuracy n the Clusterngbased scheme when σ = 2 and 4. In our smulaton, 13, 005 tranng samples were dvded nto k clusters. Accordng to our smulaton result, the Clusterng-based scheme can contnuously mprove the average postonng errors for k 60. However, for k > 60, the average postonng error almost remans the same. In our experence, for gven a total number of t tranng samples, settng k = t s a farly good choce (n the above case k 110). We wll adopt ths settng n subsequent dscussons. (2) Comparson of postonng errors: Table 1 shows the average postonng error of our scheme, as opposed to the NNSS and the deal NNSS schemes under varous combnatons of σ and η values. Note that σ reflects the mpact of envronmental nose to those schemes, whle η reflects the mpact of temperature, humdty and varous knds of hardware-dependent dynamcs. Compared to the NNSS scheme, the Clusterng-based scheme reduces the average postonng error by 16% 28% and the Regresson-based scheme reduces the average error by 35% 45%. In our experence, the Regresson-based scheme performs slghtly better than the Clusterng-based scheme. For σ = 2 and 4, Fg. 5(a) and (b) show the respectve CDFs of the postonng errors of these schemes when η s n nterval [3.1, 3.5]. The deal NNSS has maxmum error dstances of 4.5 m and 9.4 m, whereas the NNSS has maxmum error dstances of 14.8 m and 24.1 m, when σ = 2 and 4, respectvely. Under the same envronment, the maxmum error dstances of the Regresson-based scheme and the Clusterng-based scheme are approxmately m and m, respectvely. All schemes suffer from ncreased nose levels. For σ = 2 and 4, (12)

7 C.-C. Lo et al. / Pervasve and Moble Computng ( ) 7 Table 1 Average postonng errors n meters. σ = 2 σ = 4 σ = 2 σ = 4 η n [3.1, 3.5] η n [3.1, 3.5] η n [2, 4] η n [2, 4] Ideal NNSS Regresson-based Clusterng-based NNSS Fg. 5. Comparsons of CDFs of postonng errors when (a) σ = 2 and η [3.1, 3.5], (b) σ = 4 and η [3.1, 3.5], (c) σ = 2 and η [2, 4], (d) σ = 4 and η [2, 4]. Fg. 6. Correlaton coeffcent of the used rado map and the current rado map versus η when (a) σ = 2 and (b) σ = 4. Fg. 5(c) and (d) show the respectve CDFs of postonng errors usng these schemes when η s n [2, 4]. A larger nterval for the envronment- and hardware-dependent constant η makes the resultng rado map more dynamc, thereby ncreasng the dffculty and errors assocated wth of our postonng. (3) Correlaton of rado maps: To explan why the Regresson-based scheme performs slghtly better than the Clusterngbased scheme, we analyzed the correlaton between the selected rado map and the current rado map. Theoretcally, f we select a rado map wth a hgh degree of correlaton (.e., smlarty) wth the current rado map, the pattern-matchng approach should provde a postonng result wth a hgh degree of accuracy. Ths s the man dea n [9,10,21,22] and n the proposed scheme. In Fg. 6(a) and (b), we compare the correlaton coeffcents of the selected rado map wth the current rado map when σ = 2 and σ = 4, respectvely. Here, the correlaton of two rado maps s defned as follows. Let

8 8 C.-C. Lo et al. / Pervasve and Moble Computng ( ) Fg. 7. Impact of beacon densty on postonng accuracy. Table 2 Comparson of average and maxmal postonng errors (n meters). Regresson-based Clusterng-based NNSS Average error Maxmum error υ = [ υ,j ] j=1...m be the RSS vector n the selected rado map and υ = [υ,j ] j=1...m be the RSS vector n the current rado map. Also, let X be the set of [ υ,j ] =1...n,j=1...m and Y be the set of [υ,j ] =1...n,j=1...m. Then, the correlaton coeffcent ρ X,Y of two rado maps can be derved as ρ X,Y = cov(x, Y) σ X σ Y, where cov(x, Y) s the covarance between X and Y, σ X s the standard devaton of X, and σ Y s the standard devaton of Y. Here, we set η = 2 4 n the postonng phase to generate the current rado map and compare t aganst the rado maps pcked by the NNSS, the Regresson-based scheme, and the Clusterng-based scheme (η of the rado map pcked by the NNSS s 3.3). Accordng to our smulaton results, the Regresson-based scheme has a better correlaton than the current rado map, whch explans ts better performance. (4) The mpact of the densty of beacons: Fg. 7 shows the mpact of usng varous numbers of beacons. Although twelve beacons were used n the feld, we randomly selected a few to measure the sgnal strengths from nearby beacons. (The other beacons were prohbted from conductng nter-beacon measurements.) For the NNSS and the deal NNSS schemes, such a change dd not mpact performance. For the proposed scheme, usng addtonal beacons wth nter-beacon measurement capablty mproved the capture of envronmental dynamcs and thus mproved postonng accuracy. These trends are llustrated n Fg Expermental results We further verfed our results n a real envronment, as shown n Fg. 8. Note that the envronment has a dense deployment of WF access ponts (normally more than 20 40). Tranng data were collected from 124 tranng locatons, each separated by 2 m, n a publc corrdor. In each tranng locaton, we randomly collected 100 samples between July 1, 2010 and October 30, Each sample comprsed an average of ten base statons. In total, thrty fve base statons were observed. We also collected data at 117 testng locatons, each separated by 1 m, for testng purposes. In the experment, the goal was to verfy the exstence of sgnal fluctuatons and the capablty of the proposed scheme to handle such envronmental dynamcs. Note that unlke the earler smulatons, we were unable to control the values of η and σ. Therefore, we were unable to compare the deal NNSS scheme wth the proposed scheme. We randomly selected two tranng locatons and observed the measurements. Fg. 9(a) and (b) show the measured RSS dstrbutons from beacons n two locatons. Clearly, the sgnal fluctuaton problem does exst n real stuatons. Based on the above settng, we conducted a number of localzaton experments. Table 2 shows the average postonng error and the maxmum postonng error of the proposed scheme, compared wth the NNSS. Compared to Table 1, all schemes returned hgher postonng error n the real experment. However, the Clusterng-based scheme reduced the average error by 22%, and the Regresson-based scheme reduced the average error by 43%. These results verfy the effectveness of the proposed approach. (13)

9 C.-C. Lo et al. / Pervasve and Moble Computng ( ) 9 Fg. 8. Experment envronment at the Computer Scence Buldng, Natonal Chao Tung Unversty. Tranng locatons are labeled by dots ( ). Testng data were collected along the dotted lne between A and B. Fg. 9. Measured RSS dstrbutons from beacons n two locatons. 5. Conclusons In concluson, we have proposed a novel ndoor localzaton model n whch beacons measure the sgnal strengths of other beacons. Thus, beacons not only serve as localzaton tools, but also serve as calbraton tools to self-adjust the rado maps on-the-fly. We proposed two schemes to calbrate the rado map. The frst scheme s based on data clusterng, whch we call the Clusterng-based scheme, and the second scheme s based on data regresson, whch we call the Regresson-based scheme. Accordng to our results, both schemes are capable of mprovng localzaton accuracy wth the Regresson-based scheme performng slghtly better than the Clusterng-based scheme. References [1] S.J. Vaughan-Nchols, Wll moble computng s future be locaton, locaton, locaton? Computer 42 (2) (2009) [2] I. Constandache, R. Choudhury, I. Rhee, Towards moble phone localzaton wthout war-drvng, n: Proc. of IEEE INFOCOM, [3] C.-C. Lo, C.-P. Chu, Y.-C. Tseng, S.-A. Chang, L.-C. Kuo, A walkng velocty update technque for pedestran dead-reckonng applcatons, n: Proc. of IEEE Int l Symposum on Personal Indoor and Moble Rado Communcatons, PIMRC, [4] A. Kushk, K.N. Platanots, A.N. Venetsanopoulos, Intellgent dynamc rado trackng n ndoor wreless local area networks, IEEE Trans. Mob. Comput. 9 (3) (2010) [5] P. Bahl, V.N. Padmanabhan, RADAR: an n-buldng RF-based user locaton and trackng system, n: Proc. of IEEE INFOCOM, [6] T. Roos, P. Myllymäk, H. Trr, P. Mskangas, J. Sevänen, A probablstc approach to WLAN user locaton estmaton, Int. J. Wrel. Inf. Netw. 9 (3) (2002) [7] S.-P. Kuo, B.-J. Wu, W.-C. Peng, Y.-C. Tseng, Cluster-enhanced technques for pattern-matchng localzaton systems, n: Proc. of IEEE Int l Conference on Moble Ad-Hoc and Sensor Systems, MASS, [8] J. Letchner, D. Fox, A. LaMarca, Large-scale localzaton from wreless sgnal strength, n: Proc. of the Nat l Conference on Artfcal Intellgence, AAAI, 2005.

10 10 C.-C. Lo et al. / Pervasve and Moble Computng ( ) [9] L. N, Y. Lu, Y.C. Lau, A. Patl, LANDMARC: ndoor locaton sensng usng actve RFID, n: Proc. of IEEE Int l Conference on Pervasve Computng and Communcaton, PerCom, [10] J. Yn, Q. Yang, L.M. N, Learnng adaptve temporal rado maps for sgnal-strength-based locaton estmaton, IEEE Trans. Mob. Comput. 7 (7) (2008) [11] R. Sngh, L. Macch, C.S. Regazzon, K. Platanots, A statstcal modellng based locaton determnaton method usng fuson technque n WLAN, n: Proc. of Int l Workshop Wreless Ad-Hoc Networks, [12] J.J. Pan, J.T. Kwok, Q. Yang, Y. Chen, Multdmensonal vector regresson for accurate and low-cost locaton estmaton n pervasve computng, IEEE Trans. Knowl. Data Eng. 18 (9) (2006) [13] S.-P. Kuo, Y.-C. Tseng, Dscrmnant mnmzaton search for large-scale RF-based localzaton systems, IEEE Trans. Mob. Comput. 10 (2) (2011) [14] A. Kushk, K.N. Platanots, A.N. Venetsanopoulos, Kernel-based postonng n wreless local area networks, IEEE Trans. Mob. Comput. 6 (6) (2007) [15] M. Youssef, A. Agrawala, The horus WLAN locaton determnaton system, n: Proc. of ACM Int l Conference on Moble Systems, Applcatons, and Servces, MobSys, [16] A. Agwal, P. Khandpur, H. Saran, LOCATOR: locaton estmaton system for wreless LANs, n: Proc. of ACM Int l Workshop on Wreless Sensor Networks and Applcatons, WSNA, [17] M. Youssef, A. Agrawala, A. UdayaShankar, WLAN locaton determnaton va clusterng and probablty dstrbutons, n: Proc. of IEEE Int l Conference on Pervasve Computng and Communcaton, PerCom, [18] D. Fox, J. Hghtower, L. Lao, D. Schulz, G. Borrello, Bayesan flterng for locaton estmaton, IEEE Pervasve Comput. 2 (3) (2003) [19] S. Beauregard, M. Klepal Wdyawan, Indoor PDR performance enhancement usng mnmal map nformaton and partcle flters, n: Proc. of IEEE/ION Poston, Locaton and Navgaton Symposum, PLANS, [20] M. Klepal Wdyawan, S. Beauregard, A novel backtrackng partcle flter for pattern matchng ndoor localzaton, n: Proc. ACM Int l Workshop on Moble Entty Localzaton and Trackng n GPS-Less Envronments, MELT, [21] P. Krshnan, A.S. Krshnakumar, W.-H. Ju, C. Mallows, S. Ganu, A system for LEASE: locaton estmaton asssted by statonary emtters for ndoor RF wreless networks, n: Proc. of IEEE INFOCOM, [22] Y.-C. Chen, J.-R. Chang, H.-H. Chu, P. Huang, A.W. Tsu, Sensor-asssted w-f ndoor locaton system for adaptng to envronmental dynamcs, n: Proc. of ACM Int l Workshop on Modelng, Analyss and Smulaton of Wreless and Moble Systems, MSWM, [23] J. MacQueen, Some methods for classfcaton and analyss of multvarate observatons, n: Proc. of Berkeley Symposum on Mathematcal Statstcs and Probablty, [24] G. Zhou, T. He, S. Krshnamurthy, J.A. Stankovc, Impact of rado rregularty on wreless sensor networks, n: Proc. of ACM Int l Conference on Moble Systems, Applcatons, and Servces, MobSys, 2004.