TrackT: Accurate Tracking of RFID Tags with mm-level Accuracy Using First-order Taylor Series Approximation

Transcription

1 TrackT: Accurate Trackng of RFID Tags wth mm-level Accuracy Usng Frst-order Taylor Seres Approxmaton Zhongqn Wang, Nng Ye, Reza Malekan, Fu Xao, Ruchuan Wang Abstract Rado Frequency Identfcaton (RFID) technology s wdely used to acheve ndoor object trackng and postonng. Currently, many methods need to deploy a large number of reference tags beforehand and some are lmted by antennas spacng. Further, the sgnal propagaton along Non-Lne of Sght ntroduces multpath effects whch wll challenge the accuracy of RFID localzaton system. In ths work, we propose a method based on measured phase to track moble RFID tags wth mllmeter level (mm-level) accuracy. We frst partton the survellance regon nto square grds at mm-level and suppose that there s a vrtual tag as the same as the tracked one n each grd. On ths bass, for the case where the tags move along a known track wth constant speed, we only need to locate the tag s ntal poston. We leverage phase perodcty to obtan some canddates and then elmnate poston ambguty by double dfference true phase. And for the case where the tag s movng track s unknown to the system, we adopt a frst-order Taylor seres expanson to calculate the relatve dsplacements of the tracked tag and then locate the ntal poston as the same process as trackng the known trajectory. In our experment, our soluton can acheve a mean error dstance of 0.26cm and 0.55cm for known and unknown movement tracks respectvely. Index Terms Phase detecton, Rado frequency dentfcaton, RFID tags, RF sgnals. The research s support by Natonal Natural Scence Foundaton of P. R. Chna (No ), Major Program of Jangsu Hgher Educaton Insttutons (No.14KJA520002), the Key Research and Development Program of Jangsu Provnce (Socal Development Program) (No.BE ), Jangsu Planned Projects for Postdoctoral Research Funds (No C), Chna Postdoctoral Scence Foundaton (No.2014M560440), and Natonal Research Foundaton, South Afrca (AOX220). Zhongqn Wang s wth the College of Internet of thngs, Nanjng Unversty of Posts and Telecommuncatons, Nanjng, , Chna (e-mal: zhongqn.wang.cn@eee.org). Nng Ye s wth the College of Computer, Nanjng Unversty of Posts and Telecommuncatons, Nanjng, , Chna, and also wth the Jangsu Hgh Technology Research Key Laboratory for Wreless Sensor Networks, Nanjng , Chna (e-mal: Reza Malekan s wth the Department of Electrcal, Electronc and Computer Engneerng, Unversty of Pretora, Pretora, South Afrca (e-mal: reza.malekan@up.ac.za). Fu Xao and Ruchuan Wang are wth the College of Computer, Nanjng Unversty of Posts and Telecommuncatons, Nanjng, , Chna, and also wth the Jangsu Hgh Technology Research Key Laboratory for Wreless Sensor Networks, Nanjng , Chna. He s also wth Key Lab of Broadband Wreless Communcaton and Sensor Network Technology of Mnstry of Educaton, Nanjng Unversty of Posts and Telecommuncatons, Nanjng, , Chna (e-mal: {xaof, R I. INTRODUCTION ADIO Frequency Identfcaton (RFID) technology s ncreasngly used n varous applcatons such as asssted trackng of robots [1], product dentfcaton[2], asset assessment, ndoor postonng and so on. Currently, the common method to locate passve tags n practce s descrbed as follows. People often deploy many RFID readers n dfferent montorng areas to contnuously read RFID tags. The tags are assgned wth the unque Electronc Product Code (EPC) and gven poston nformaton beforehand n the database. Once a reader captures a new tag s EPC, people consder that the tag has been moved to the reader s survellance area. However, there are many dsadvantages n ths coarse method: () Low postonng resoluton. The readng range of an RFID reader antenna s generally about 3~10 meters, and RFID readers can only provde absence and presence results, so the system s far from meetng the hgh accuracy requrements. () False negatve reads [3], [4] and false postve reads [5], [6]. The former means that a reader fals to read a tag n the readng zone and the latter means that a tag n some other areas outsde the ntended read zone s read. These two problems can also affect the postonng accuracy. Many applcatons wll beneft from mllmeter level (mm-level) localzaton accuracy. For example, false postve reads wll be avoded by settng the ntended readng zone beforehand. If the RFID tag s wthn the area, the RFID reader wll record and report events related to ths tracked tag. If not, the reader wll gnore t. As another example, n a large-scale clothes shop a retaler could use the RFID locaton system to vsually track clothes wth RFID tags, makng sure that sales representatves could easly fnd matchng clothes whch may have gone astray and customers could easly know the locatons of wanted clothes. At present, the two key approaches for RFID localzaton are based on receved sgnal strength ndcaton (RSSI) and rado frequency (RF) phase. () RSSI. The RSSI methods [7]-[11] need to deploy many reference tags snce the absolute calbraton for RSSI measurements s rather dffcult, and postonng accuracy s greatly affected by antenna desgn, mpedance matchng, and the changes n reflecton coeffcent [12]. The dstance error s about 60cm. At present, many commercal-off-the-shelf (COTS) RFID readers can report RF phase once an RFID tag s

2 2 3 successfully read. Phase resoluton can reach radans, offerng ( ) / cm rangng resoluton for an RF carrer wave at the frequency of MHz. In addton, the RSSI-based methods are not a good choce because the propagaton envronment wll more easly affect the RSSI measurements than phase [13]. As a result, the method to track tags movement trajectores based on RF phase wth hgher resoluton and better nose-tolerant ablty than RSSI has receved many researchers attenton. () RF Phase. Angle of arrval (AoA) [14]-[16] uses multple antennas to receve the tag s phase and then computes ther angles based on phase dfference. But these methods need to put a strct constrant on the antennas spacng. Backpos [17] deploys more than three antennas to read tags phase, where two adjacent antennas should be wthn a spacng of half a wavelength. After that, they use the hyperbolc postonng method to locate tags. However, f adjacent antennas spacng s greater than half a wavelength, the method wll fal to deal wth the phase perodcty, leadng to numerous possble canddates. PnIt [18] captures and extracts multpath profles of the target tag and reference tags at known postons va an antenna moton, and then adopts dynamc tme warpng (DTW) technques to pnpont a tag s locaton. However, t needs to deploy dense reference tags n advance and can t track moble tags. DAH (Dfferental Augmented Hologram) [19] can track known and unknown trajectores of moble tags wth a mean error dstance of 0.6cm and 12cm respectvely usng COTS RFID readers n a lab envronment. The man dea s to construct a vrtual antenna array n tag movement and leverage the statstcal method to fnd the optmal trajectory. DAH s effectve way to track and locate movng tags than prevous work. However, thermal nose and external nference, perhaps from human movng besdes RFID tags n actual crcumstance wll both affect the reported RF phase. Also, as the survellance area expands, the computatons wll jeopardze the real-tme performance. In addton, the accuracy of trackng unpredctable movements should be mproved n some extent. In ths paper, we propose a novel postonng and trackng method based on measured phase usng COTS RFID devces, called TrackT. The phase perodcty s manly used n our approach to mprove the ablty of nose tolerance and real-tme performance n trackng moble tags. At frst glance, the phase perodcty s a negatve factor and many researchers try to elmnate the unknown parameter n prevous work. However, the dfferences of the phase perodcty between two contnuous reads are an effectve ndcator ntroduced n the paper, wth stronger nose-tolerance than phase-dfference method lke DAH. The basc dea of TrackT s descrbed as follows. We frst dvde the survellance regon nto mm-level square grds and assume that there s a vsual tag as same as the tracked tag on the center of every grd. () Known Movement. We compare the sequence of the dfference of phase perodcty between each vrtual tag and the physcal one to acqure several canddate ponts and then separately compute the correspondng double dfference of true phase to fnd the mnmum value. So the grd wth the mnmum value s chosen as the optmal poston of the tracked tag. Once the ntal poston has been determned, the predefned track wll also be ftted. () Unknown Movement. In ths case, both the ntal poston and the tag s dsplacements should be estmated. We suppose that all of antennas read the tag at the same tme, and then the tag s dsplacement every read round could be calculated usng frst-order Taylor seres expanson. For the unknown ntal poston, we leverage the same method of trackng known movement trajectory to select Fg. 1. Overvew of RFID system. It shows a conceptual dagram of the rado wave propagaton between an RFID reader and a passve RFID tag. the optmal one. For achevng real-tme performance, a rule to reduce the computatons s also proposed n the paper. Summary of Results: We buld the system usng Impnj R420 RFID reader, four 8dB antennas and RFID passve tags. Our man results are descrbed as follows: () When four RFID antennas are deployed around the montorng area, TrackT can acheve mm-level accuracy wth mean error dstances of 0.26cm and 0.55cm for controllable and uncontrollable movements. () TrackT s a real-tme localzaton system so the computaton tme for producng an ntermedate result should be less than the tme nterval of about 33ms between successve nventores of the same tag. After optmzaton, TrackT can acheve a mean of about 11.21ms computaton tme, whch can meet the real-tme requrement. Contrbutons: TrackT can capture the known and unknown movement trajectores of moble RFID tags wth mm-level accuracy. As far as we know, TrackT s the frst to leverage the phase perodcty wth hgh nose tolerance to track moble tags. Besdes, we also explot the localzaton theory of carrer phase measurement n global postonng system (GPS) to track unknown trajectory of the movng RFID tag. As a result, TrackT wth less postonng error and hgher real-tme performance can be easly mplemented n real applcatons based on COTS RFID devces. The remander of the paper s descrbed as follows. The background and emprcal studes are ntroduced n Secton II. The man desgn of the TrackT approach s presented n Secton III. We dscuss addtonal detals n Secton IV. The mplementaton and evaluaton are gven n Secton V. Fnally, Secton VI ntroduces the lmtatons and concludes the paper. II. PRELIMINARIES In ths secton, we manly ntroduce RF phase, Doppler frequency shft, sesson persstence and phase perodcty.

3 3 A. RF Phase For an RF carrer wave at frequency f (Hz), the relaton between frequency f and wavelength s gven by c (meters) (1) f where n ar, the speed c of the electromagnetc wave s equal to the speed of lght,.e., m/s. In Chna, a typcal ultra-hgh frequency (UHF) reader has 16 channels workng at MHz ndustral scentfc medcal (ISM) band. As shown n Fg. 1, the dstance between an RFID reader antenna and a passve RFID tag s R. In addton to the RF phase over dstance, RFID hardware nstruments ncludng the reader s transmt crcuts, the tag s reflecton characterstc, and the reader s recever crcuts wll all ntroduce some addtonal phase shfts, T and Tag respectvely [12]. So the true R phase can be expressed as 2R N 2 T R (2) Tag where, the output parameter from the RFID reader, s the measured phase wthn 0,2. The unknown parameter N, an ntegral multple of 2, s called phase perodcty. B. Doppler Frequency Shft Doppler frequency shft s the shft n frequency of the receved sgnal at the reader due to relatve moton between the reader and the tag. Assume that the carrer frequency s f and the sgnal frequency caused by the Doppler effect s f, so the r Doppler frequency shft s defned as follows: 2 f d f fr (3) c dt where d s the radal velocty wth whch the tracked tag dt moves away from or approaches the antenna. The factor of 2 arses from the tag backscatterng the reader s carrer sgnal. So the phase shft n caused by Doppler effect s: tt n 2 t f fr dt t T 2 f d (4) 2 t dt c dt 4 tt t where t and T are the dstances from the tracked tag to t the antenna at the tme tt and t respectvely. T denotes the tme duraton of a packet related to EPC length and reverses lnk rate. If tt, the tag moves away from the antenna, t.e., n 0. If tt, the tag moves close to the antenna, t.e., n 0. Ths feature s often used to determne tags n moton versus statonary tags and tag moton drectonalty. Consderng Doppler effect, the true phase s denoted by N n. In COTS RFID readers, the slower reader nventory modes (e.g. Dense Reader, M=8) should be confgured to measure Doppler frequency shft because longer measurement ntervals typcally provde more accurate Doppler frequency shft estmates. But n ths paper, the reader mode wth hgher read rate (e.g. Max Throughput) s used n our proposed method. In ths case, the accuracy of the reported Doppler frequency shfts wll be reduced sgnfcantly. Fortunately, snce T s generally at mllsecond level for the low-speed RFID tags, t s reasonable for us to gnore the mpact of Doppler effect on the true phase n the followng. C. Sesson Persstence There are two nventory flags of A and B for a tag. All nventory flags default to A when a tag powers up and the flag B s the state when tag power s lost. Regardless of the state of A and B, tags both allow nventory operatons. The persstence of the state B s related to the nventory flag confgured by RFID reader and tag power state. Indeed, there are four search modes n RFID reader,.e., Dual Target Sesson 0 (Persstence tme T S =0s), Sngle Target Sesson 1 (T S =0.5~5s), Sngle Target Sesson 2 or 3 (T S =2~60s) and Sngle Target wth Suppresson (T S =0.5~5s). For Dual Target mode, the state of the Sesson 0 flag perssts ndefntely when tag power s on, but the state of the flag s lost mmedately when tag power s lost. Ths sesson mode enables the reader to read tags contnuously, whch s sutable for read small batch of products n statc envronment. And Sngle Target mode has a fxed and lmted persstence even when tag power s on; f the flag has not been refreshed by an nventory operaton n more than the persstence tme. Ths model s sutable for read large products n dynamc envronment [20]. Due to the nfluence of multpath propagaton and tag populaton sze, the persstence of state B s uncertan for RFID reader, meanng that the tme between successve nventores of the same tag s random. In summary, a tag can be nterrogated for about 30 tmes per second under the read mode of the Sesson 0 flag [21], so the search mode of Dual Target Sesson 0 s employed n our proposed method. D. Phase Perodcty Suppose that two successve tme-phase pars are measured for the same tag,.e., t, and t1, 1. In ths case, the true phase are N and 1 1 N, and the 1 Eucldean dstances between the RFID tag and the antenna are and respectvely. We also assume that 1 d represents the tag movement dstance between t and t. Snce the 1 system suffers from serous packet loss when the tag moves wth a hgh speed, resultng that the tag even cannot be nterrogated [22], we consder the case where tags move wth a low speed n our system. In addton, the nterval between successve nventores of the same tag s short enough, so t s feasble to assume d /4. Based on the trangle constrant, we have d,.e., Snce N -N 1 1 (5) N s an ntegral multple of 2, the dfference of

4 4 phase perodcty between two contnuous reads, denoted as N, can be 0 1 N N 1- N (6) Note that the tme between successve nventores of the same tag wll depend on reader mode, tag populaton sze, and envronmental condtons (e.g. nterference levels). Whle the RFID reader mode s set to the hghest read rate, the mnmum update nterval s about 0.5 seconds for the populaton of 512 RFID tags n the read area. The tag populaton sze n the read range should be consdered to make sure d /4 n practce. Many methods [23],[24] are proposed to mprove the tme-effcency for large-scale RFID systems. III. TRACKT OVERVIEW In ths secton, we gve detals on our trackng approach n two cases. In 2D scenaro, a tracked RFID tag s algned wth the center of antennas. We frst ntroduce how to track a moble RFID tag s trajectory n the case of conveyor wth known track that the trajectory functon and the tag s speed are known for us before trackng. Then we descrbe how to track the unpredctable movement that both trajectory functon and speed are unknown n pror. A. Movement wth known track Assume that the phase sequence read by an RFID reader n tag s movement s { 1, 2, 3,..., n }. Whle the tag moves along dfferent trajectores n the montorng area, the sequence s also dfferent from each other. In ths case, the tag movement trajectory could be captured by dstngushng the correspondng phase sequence. However, t s very dffcult to collect all of phase sequence generated by dfferent trajectores for a tag movng n the regon. In addton, the reported phase s suffered from the effect of the hardware nose and external nterference, and the tme between the two contnuous nventores s random even on the same movement trajectory. In summary, there are above lmtatons to capture the tag s movement trajectory by comparng dfferent phase sequences. The sequence of the dfference of phase perodcty s N { N1, N2, N3,..., N n -1} correspondng to based on (6). In our experment, the measured phase follows a typcal Gaussan dstrbuton wth a standard devaton of 0.1 radans, N 0, 0.1 2, but the slght then 1 ~ varaton of phase estmates has lttle mpact on N. For example, due to 1.2 radans and 2.5 radans, ther correspondng N are both equal to 0. So we consder that the dfference of phase perodcty s not senstve to nose nterference. Moreover, unlke the phase values varyng wthn 0,2, N has only three certan values,.e., 0 and 2. Therefore, the sequence of the dfference of phase perodcty s employed to track the tag s trajectory and locate the ntal poston n ths paper. 1) Preprocessng In the 2D scenaro of RFID tag s movement wth known track, the tracked target attached on an RFID tag moves wth a constant speed. RFID antennas are deployed around the survellance regon to read the tag for trackng and postonng. At present, the most common unform tracks are a varety of conveyor belts, ncludng unform lnear and crcular track. Conveyors are wdely used for movng and sortng goods packaged n cartons, boxes, or shrnk-wrapped contaners. If we can know the ntal poston on the track, the followng trajectory can be determned. Based on ths observaton, we frst make the followng assumptons: () Suppose that the tag wth unform moton n the survellance regon keeps n the statonary state, so each of fxed antennas unformly moves along the opposte drecton relatve to the tag. () The survellance regon s dvded nto W L square grds at mm-level (less than 1cm) and suppose that there s a vrtual tag as the same as the tracked one on the center of each grd. 2) Comparson rule In the rectangular coordnate system, let the ntal coordnate X T, Y T. Suppose that there are m reader of the tag be 1 1 antennas deployed around the survellance regon to read the tracked tag. Once recevng n reads for each antenna, TrackT wll calculate the ntal poston of the movng tag. The j th read tme and phase value measured by the th antenna are t A, j and A, j respectvely. The dfference of phase perodcty s denoted by N A, j. In addton, as the tracked tag moves, we assume that reader antennas move along the opposte drecton, so the vrtual poston of the th antenna s A A t A -t A, where represents the,j,1,j,1 speed functon of the moble tag and A s the coordnate of,1 the th fxed antenna,.e., A,1 X A,1, Y A,1 Defnton 1 G, WL, A, j. s the j th theoretcal phase value of the vrtual tag on the grd G read by the th antenna. W, L Ignorng the addtonal phase, T and Tag, the theoretcal R phase s defned as follows: 4 GW, L, A, j GW, L A, j mod 2 (7) where G A s the Eucldean dstance between the grd WL,, j G and the vrtual antenna W, L A. Also, the symbol mod, j means the modulo operaton because the reported phase s a perodc functon wthn 0,2. N G A s the dfference value of phase Defnton 2, WL,, j perodcty between G, WL, A and, j G, WL, A, j1 :

5 5 0 GW, L, A, j1 GW, L, A, j W, L, j W, L, j1 W, L, j 2 2 GW, L, A, j1 GW, L, A, j Defnton 3 WL, N G, A 2 G, A G, A 2 (8) Num G s called the poston correlaton between the vrtual tag on the grd G and the tracked tag, W, L whch s defned as follows: to the tracked tag. The correspondng coordnate s (49.75,29.75), devatng from the ground truth about cm. However, n our experment we fnd that there may be multple canddates as the number of collected phase data decreases. In Num G only usng the phase from Fg. 4, f we calculate WL, antennas 1 and 2, we have NumG Num G. 100,60 65, The correspondng coordnates are respectvely (49.75,29.75) and (32.25,26.25), and the dstance between the two coordnates s 17.85cm. So we beleve that phase perodcty results n multple maxmums n the survellance regon. Fg. 2. Survellance regon. The survellance regon of 80 cm80 cm s dvded nto the number of square grds wth the wdth of 0.5 cm and then mapped nto the coordnate system. The XY-axs unts are both centmeters. The antennas 1-4 are respectvely deployed on (0,0), (110,0), (100,110) and (0,110). The tracked tag moves along a known track from the ntal poston of (50,30). m n1 W, L, j, W, L,, j Num G f N A N G A (9) 1 j1 where m s the number of antennas, n s the read tmes n a read round for each antenna and f N A,j, N GWL,, A,j s the poston correlaton functon between N(A, j) and N( G, A ) : WL,,j, j WL,, j, j WL,, j 1 N A N G, A f N A, j, N GWL,, A, j (10) 0 N A N G, A The smaller the dstance of the tracked tag to the vrtual tag on the grd s, the larger the value of Num s. In ths case, W,L the grd wth the maxmum WL, Num G could be regarded as the optmal poston of the tracked tag. To vsually understand ths dea, let us consder the smple example shown n Fg. 2. The range of survellance regon s 80cm 80cm composed by square grds wth the wdth of 0.5cm. The RFID reader antennas 1-4 are deployed on both sdes of survellance regon, where the coordnates of antennas 1-4 are respectvely (0,0), (110,0), (100,110) and (0,110) n unt of centmeter. The ntal poston of tag before the tag movement s (50,30), movng along the negatve drecton of x-axs wth the speed of about 20cm/s. We collect the about 300 phase values by all of antennas n 2.5 seconds, so the expected maxmum of Num wll research 296. In Fg. W,L 3, the maxmum pont on the grd W=100, L=60 s 286, meanng that the poston of the vrtual tag on the grd s closest Fg. 3. The optmal poston on the grd. The tracked tag locates potentally on the grd of W=100, L=60 wth the maxmum poston correlaton. Fg. 4. Poston ambguty. Wth the decrease of collected phase data, there are two grds wth maxmum values, resultng n the poston ambguty of tracked tag. 3) Dfferental true phase for poston ambguty In the followng, we ntroduce how to elmnate the poston ambguty based on the true phase defned n Secton II. G A s the j th true phase of the vrtual Defnton 4, WL,, j tag on the grd G read by the th antenna. Here, we also W, L gnore, T and Tag, so the true phase wth phase R perodcty s denoted as follows: 4 G, A G A (11) W, L, j W, L, j Defnton 5 A,j s the j th true phase of the tracked tag read by the th antenna: A A N A (12),j,j,j where A, j s the reported phase value and, j N A s the phase perodcty of the tracked tag. s called Sngle Dfference, whch Defnton 6 G WL, s calculated by

6 6 m n GW, L GW, L, A, j A, j m n 1 j1 1 j1 GW, L, A, j N GW, L, A, j A, j N A, j (13) where N A, j s an unknown parameter for us and N( G, A ) s the phase perodcty of vrtual tag on the grd G W,L W, L, j. Based on the known dstance G A W, L. j, we have 4 N GW, L, A, j GW, L A. j \ 2 2, where the symbol \ represents the quotent operator to obtan quotent of 4 GW, L A and 2.. j Obvously, the closer the vrtual tag on the grd s to the ntal poston of the tracked tag, the smaller the value of s. Next we wll take dfference of the sngle dfference W,L to elmnate the unknown parameter N. A, j that k 1, although the value of N A, j N A,1 be the maxmum. Then N A, j N A,1 as: wll not can be expressed, j N A,1 N A, j N A, j1 N A, j1 N A, j2... N A,2 N A,1 N A (17) Fg. 5. The results of double dfference. Compared wth the values of double dfference on the grds of W=100, L=60 and W=65, L=53 wth same poston correlaton n Fg. 4, the optmal poston s on the grd of W=100, L=60 wth the mnmum value. And the postonng result of TrackT on the grd of W=100, L=60 has a lower error dstance than that of DAH on the grd of W=100, L=62. Defnton 7 DffGWL, 1 j1 s called Double Dfference, whch s calculated by: m n1 Dff GW, L GW, L, A, j A, j GW, L, A,1 A,1 (14) Hence, the ntal coordnate can be calculated by the functon f X T1, Y T1 mn Dff GWL,. And double dfference can reduce the effect of measurement nose as much as possble, whch can be proved as follows: For the j th and k th phase values collected by the th antenna, we have GW, L, A, j A, j GW, L, A,k A,k GW, L, A, j GW, L, A,k A, j A,k (15) 4 GW, L A, j GW, L A,k A, j A,k N A, j N A,k For the unform lnear track, as the tracked tag moves away from the ntal poston, we have: max N A N A N A N A (16),j,k,j,1 Note that for the other moton trajectores, we also assume Fg. 6. Curvlnear moton. The RFID tag moves wth unpredctable movement n the montorng area. Based on (6), we depend on the collected phase, {,,..., }, to calculate the every dfference value of A,1 A,2 A,j, j N A,1. Because all of A j N A, j 1,2,... are ndependent standard normal random varables wth the same dstrbuton, N 0, for the th antenna, we have.e., A, j ~ A, j A,1 ~ N(0, 2 ). As descrbed before, 0.1radans, so the standard devaton of phase dfference N A N A k, the s 0.14 radans. However, snce,j,1 2 value wll be far larger than 0.14 radans. In ths case, the varance of phase values caused by thermal nose and external nterference has lttle mpact on the fnal result of Dff G for dfferent antennas. In ths way, TrackT has WL, strong nose tolerance to multpath effect n practce. Then we put all of canddates wth the same poston correlaton nto (14) to calculate the results of double dfference, so the grd wth the mnmum value s consdered as the fnal poston of the tracked tag. In Fg. 4, there are two canddates on the grd of W=100,L=60 and W=65,L=53 and the correspondng values of double dfference are respectvely Dff G Dff G The 100,60 and 65,53 optmal poston s obvously on the grd of W=100, L=60 and the correspondng coordnate s 49.75, The dstance error devates from the ground truth about cm. In addton, there s a mnmum value on the grd of W=100, L=62 n the survellance zone. The correspondng double dfference Dff G 11.28, and the dstance error s s 100, cm. The postonng accuracy n TrackT has a lower

7 7 error dstance than that n DAH [19] based on only double dfference, whch wll be verfed n the Secton V. B. Movement wth unknown track As shown n Fg. 6, an RFID tag moves along the curvlnear track n 2D scenaro. Smlarly to the prevous method of trackng the known trajectory, we also assume that each of antennas moves n the opposte drecton of the tag movement. The tme nterval between the j th and (j+1) th nventores of the tracked tag read by the th antenna s t A, j, so durng the tme nterval, horzontal and vertcal dsplacements of the th antenna relatve to the tag can be respectvely represented as x A, j y A, j, whch s subject to and 2 2 max x A, j y A, j. 4 RFID employs a slotted-aloha meda access scheme, whch means that the order n whch tags are nventored wll be random [20],[25], so antennas can t read the tag at the same tme. Suppose that n every round each antenna reads the tag one tme. The Eucldean dstance of the tracked tag to the th antenna s A, j at the tme t A, j. Based on the relatonshp between the true phase A, j and the dstance of the tag to th antenna, we have A A N A 4, j, j, j 1, j 1, j 2 2 X T X A Y T Y A X T, Y T s the unknown ntal coordnate of where 1 1 (17) the tracked tag n the curvlnear moton and X A, j, Y A, j s the coordnate of the th vrtual antenna at the tme t A, j. A frst-order Taylor seres approxmaton can make unsolvable problems possble for a restrcted doman [26]. So we compute a frst-order Taylor seres expanson around the, : pont X A, j Y A, j of the functon A, j 2 2 A, j X T1 X A, j Y T1 Y A, j X T1 X A, j1 Y T1 Y A, j1 A, j1 A, j1 A x A y A, j1, j1, j1 (18) where the coordnate pont of the th vrtual antenna at the tme t A can be denoted by, j X A, j X A, j1 x A, j1 Y A, j Y A, j1 y A, j1 For A, j 1 and A, j, we have (19) A, j A, j1 4 A, j A, j1 N A, j N A, j1 2 2 X T1 X A, j1 Y T1 Y A, j1 A, j1 1, j1 1, j1 x A, j1 y A, j1 A, j1 A, j1 X T X A Y T Y A Thus, A A 4 N A N A, j, j 1, j, j1 j A j j A j X T X A Y T Y A 1, 1 1, 1 x A, j1 y A, j1, 1, 1 (20) (21) In our experment, we deploy 4 drectonal antennas to acqure phase data for trackng a moble tag accurately. Snce n a round the read tme among dfferent antennas are very close to each other, about 33ms, t s reasonable for us to assume that each antenna reads the measured phase at the same tme, whch means the tag dsplacements measured by 4 antennas between successve read rounds are the same,.e., x A x A y A y A. Based on, j1 j1 and, j1 j1 phase data collected by 4 antennas, the equaton (21) can be denoted n matrx form,.e., Ax B (22) where 1 1, j1 1 1, j1 A1, j1 A1, j1 1 2, j1 1 2, j1, A2, j1 A2, j1 1 3, 1 1 3, 1 j j A3, j 1 A3, j 1 1 4, j1 1 4, j1 A4, j1 A4, j1 A1, j A1, j1 N A1, j N A1, j1 A2, j A2, j1 N A2, j N A2, j1 A3, j A3, j1 N A3, j N A3, j1 A4, j A4, j1 N A4, j N A4, j1 [ T j1 j1 ], j 2 X T X A Y T Y A X T X A Y T Y A A X T X A Y T Y A X T X A Y T Y A B 4 x x A y A The coordnate (0, 0) s frstly chosen as the ntal pont of the tracked tag. We can estmate the dsplacement x of vrtual antennas usng least squares [27], [28]. After enough dsplacements are generated, the unknown ntal coordnate X T, Y T can be calculated by acqurng the maxmum 1 1 values of the poston correlaton Num G WL,, and elmnatng the poston ambguty as the same process as trackng the known trajectory of a moble tag. In addton, snce the vrtual antennas move n the opposte drecton to the tracked tag, the horzontal and vertcal dsplacements of the tag n practce are

8 8 x T j1 x Aj 1 and ytj 1 y Aj 1. So a completed trajectory of unknown movement s ftted from the above steps. IV. DISCUSSION In ths secton, we attempt to answer some practcal ssues. gnored because the values of WL, equal to zero. And the maxmum value of WL, Num G for many grds are Num G s equal to 202 on the grd of W=101,L=60. The correspondng coordnate s (50.25, 29.75) and the dstance error devates from the ground truth about cm. As a result, the optmzaton method to reduce the computatons wll be employed n the followng experment. Fg.7. The rule of reducng computatons. If the tracked tag maybe on the grd, G, A Tag A should the dstance dfference between W, L and, be less than A. How to reduce the computatons n TrackT? We know that the read range of a reader s antenna s generally about 10m, so the survellance regon composed of multple antennas s so large enough that the number of grds at mm-level n the regon s also huge, whch wll jeopardze the real-tme trackng capablty. From Fg. 3 and 4, the value of the deeper blue grds s smaller than others. That s to say, the ground truth of tracked tag doesn t exst on these grds so the related computaton s unnecessary for us. As shown n Fg. 7, assume that the grd wdth s a, the dstance from the center of grd to four angles s 2 2 a and the dstances from the grd centrod and the poston of tracked tag to the th antenna are respectvely G W, L, A and Tag A. If the tracked tag s on the grd, then, 2 G W, L, A Tag, A a. The measured phase of 2 the tracked tag by the th antenna. Ignorng the A s phase perodcty, the phase of the vrtual tag s 4 GW, L, A ( GWL,, A) mod 2. If the tracked tag s on the grd W, L 2 2 a. G, then (, ) 2 GW, L A A a 4. So we 2 Num G of only need to compute the poston correlaton WL, the grd meetng the above lmtaton. Our evaluaton shows that 65% of computaton wll be reduced usng ths method. In addton, the lmtaton wll not affect the postonng accuracy. For example, n Fg. 3 we calculate Num G WL, between vrtual tags on the every grd and the tracked tag. Here, we use above lmtaton to reduce computatons and the result s shown n Fg. 8. A majorty of non-canddate grds can be A Fg. 8. The optmzed result Compared wth Fg. 3, the postonng results before and after optmzaton have the same dstance error devated from the ground truth. Fg. 9. 3D space. The tag and the center of antenna aren t on the same plane. In ths case, the survellance plane s composed of cubes wth mm-level wdth, whch wll be nevtable to ncrease computatons. B. How to work wth Multpath effect and Non-Lne of Sght? Multpath effect s a common ssue for wreless localzaton systems. Undesred sgnals n the actual envronment can combne wth the prmary backscatter, thereby ncreasng or decreasng the receved sgnal power at the antenna recever. The sgnal propagaton along Non-Lne of Sght (NLOS) ntroduces multpath effects to challenge the accuracy n phase measurements. TrackT works wth multpath effect from three aspects. At frst, the dfference values of phase perodcty wth strong nose tolerance are manly leveraged n TrackT to locate RFID tags. TrackT fnds the optmal trajectory of the moble tag by multple reads. Second, the tag s moblty can help antennas read phase data from dfferent drectons. Lots of measurements from varous drectons can reduce the nfluence of multpath effect. Thrd, there are generally 16 frequency channels for an RFID reader. We can select the most senstve frequency to the tag n order to reduce the mpact of multpath propagaton. Currently, there have been many excellent solutons [29], [30] n recent years. Ther basc deas are to fnd out the drect path among all paths. C. How to handle wth 3D scenaro? In practce, t s mpossble that all of tags are algned wth the

9 9 center of antennas. In Fg. 9, the dstance from the tag to the center of antenna s larger than. If we suppose that they are on the same plane, t wll be nevtable to ncrease postonng error. In order to obtan the accurate trajectory, we can extend our method to 3D scenaro. In ths case, the survellance plane s composed of cubes wth mm-level wdth. However, the computaton of trackng tags n 3D space s greatly huger than that n 2D scenaro, whch wll reduce real-tme performance. So we wll take further study n our future work. Fg. 10 (a). Lnear Track deployed on both sdes of the track. In the rectangular coordnate system, the coordnates of the antennas are (0,-60), (100,-60), (100,50) and (0,50) respectvely, as shown n Fg. 10 (a). The center of antennas perpendcular to the track s assgned wth tags attached on the toy tran. The Eucldean dstances from antennas 1, 2 and 3, 4 to the track are respectvely 60cm and 50cm.The length of lnear track s 80cm. Then we start the toy tran wth the speed of 20cm/s. Ignorng the non-unform moton state at the begnnng, the ntal poston of the frst tag s (40,0). Crcular Track: Antennas 1, 2 and 3, 4 are respectvely deployed around the survellance regon and the coordnates of the antennas are (-150,0), (0,-150), (150,0) and (0,150) respectvely, as shown n Fg. 10 (b). The center of antennas s also assgned wth tags. The radus of the crcular track s 50cm. The ntal poston of the tag s (50,0). In the followng experment, the grd wdth of the survellance regon s set to 0.25cm. All of experments are mplemented n a relatvely open ndoor envronment to reduce the nfluence of multpath propagaton and envronmental nterference as much as possble. 2) Software We depend on LLRP Toolkt programmers gude [19] to wrte RFID applcatons communcatng wth the R420 RFID reader n C#. The software s shown n Fg. 11. The collected RF phase estmates are handled n Matlab R2010a to evaluate the TrackT performance. We run the software at a Lenovo Thnkpad X220 computer equpped wth dual processor CPU of Core M processors at 2.4 GHz. Fg. 10(b). Crcular Track V. IMPLEMENTATION AND EVALUATION In ths secton, we present the mplementaton and conduct performance evaluaton on the prototype. A. Implementaton 1) Hardware We mplement our method usng an Impnj R420 RFID reader, four 8dB drectonal antennas and Impnj H47 RFID tags. The RFID reader s confgured as the fastest RF mode (Max Throughput, Dual target and Sesson 0) to mprove read rate. The antennas wth the sze of 280mm 280mm 40mm are connected to the Impnj R420 reader. In Chna, UHF RFID operatng frequency band s MHz MHz wth 500 khz channel spacng. Here, the fxed frequency, MHz, s selected n our experments. A plastc toy tran attached wth Impnj H47 tags moves wth the constant speed of 20cm/s on the track. Two knds of moton trajectores are deployed n our experment- Lnear Track and Crcular Track. Lnear Track: Antennas 1, 2 and 3, 4 are respectvely Fg. 11. Software Interface. B. Evaluaton for known track 1) Error dstance between TrackT and DAH We use the scenaro of lnear track to evaluate the error dstance between TrackT and DAH. The error dstance s defned as the Eucldean dstance between the ground truth and measured poston. In the experment, the mean of tme between the two successve nventores of the same tag for an RFID antenna s about 0.033s. Snce the tag moves along the track at the speed of 20 cm/s, the mean dsplacement s about 0.66cm, far less than /4. We collect the reported phase from four antennas wthn 2.5seconds. The experment repeats for 50 tmes respectvely. The cumulatve dstrbuton functon (CDF) of poston error based on TrackT and DAH are descrbed n Fg. 12 and 13. TrackT has a mean error dstance of 0.18cm, 0.16cm and 0.26cm n x-axs, y-axs and combned dmenson,

10 10 outperformng the DAH method. Most of combned error dstances of TrackT are less than 0.5cm. So TrackT can sgnfcantly mprove the accuracy of RFID postonng. 2) Impact of tag number In the experment, we attach 3-10 tags respectvely on the toy effectveness. The experment repeats for 50 tmes respectvely wth optmzaton and wthout optmzaton. Suppose that the computaton tme of outputtng trackng result after 70 reads s t, so the mean computaton tme between the two successve nventores s t/69. Fg. 12. Trackng n lnear moton of RFID tags usng TrackT. Fg. 14. Impact of tag number. As the number of tracked tags ncreases, the postonng accuracy wll be reduced. Fg. 13. Trackng n lnear moton of RFID tags usng DAH. tran to evaluate the mpact of tag number. The tme to collect phase data s 2.5s. As the number of RFID tags ncreases, the mean of tme between two successve nventores of the same tag for an RFID antenna wll ncrease accordngly. In ths case, the total number of captured phase data for each tag wll decrease wthn the nterval of 2.5s. From Fg. 14, when there are 3 tags attached on the toy tran, the mean of combned error dstance s 0.38cm wth the standard devaton of 0.15cm. However, when there are 10 tags on the toy tran, TrackT obtans a mean combned error dstance of 0.72cm wth the standard devaton of 0.31cm. Therefore, as the number of tags n the survellance regon ncreases, TrackT should take more tme to receve enough reads for keepng the hgh postonng accuracy. 3) Realtme performance TrackT s a real-tme trackng system. Once the RFID reader collects a measured phase, TrackT wll compute the phase dfference between the current phase and the former one. The fnal trackng result wll be released after enough reads have been receved n the system. As depcted n Fg.15, the accuracy tends towards stablty when the total number of reads from all of antennas s more than 70 tmes. In addton, f there s only one tag n the survellance regon, the tme between successve nventores of the same tag s 33ms as descrbed before. In ths case, the mean of tme nterval for each calculaton should be less than 33ms n order to mantan the realtme performance of TrackT. In Secton IV, we have dscussed how to reduce the computatons, so the method s conducted here to evaluate ts Fg. 15. Stablty for known track. The accuracy tends towards stablty when the sum of reads from all of antennas s more than 70 tmes. Fg. 16. Real-tme Performance. After optmzaton, a mean tme of each calculaton s about ms wth a standard devaton ms, whch can meet real-tme performance. The expermental results are shown n Fg.16. A mean tme of each calculaton s 72.81ms and a standard devaton s 10.06ms wthout optmzaton, whch far exceeds than 33ms. The generaton wth optmzaton keeps a low tme cost, achevng a mean tme of 11.21ms wth a standard devaton 0.58ms. So the optmzaton method descrbed n Secton IV s effectve to make the system meet the real-tme requrements. 4) Impact of antenna poston At frst, we deploy antennas 1-4 on both sdes of the crcular track wth the ntal coordnates of (0,-150), (100,-150), (100,150) and (0,150). As the vertcal dstance of each antenna to the center of the track vares from 1.5m to 6m, the mpact of antenna poston on trackng accuracy s shown n Fg.17. Wth

11 11 the ncrease of the dstance to the track center, the postonng combned error wll also ncrease, and the error manly comes from the y-axs rather than x-axs. However, when we deploy antennas around the crcular track and repeat the above experment, we fnd that the combned error s unrelated to uncertanty regon (the ntersecton of the crcular lnes) s small (low DOP). In Fg. 18, however, the antennas 1, 2 or 3, 4 are very close to each other, resultng n a consderably large uncertanty regon, so the postonng accuracy wll be dluted n some extent. Fg. 17. Impact of antenna poston. Whle antennas are deployed on both sdes of the track, the localzaton error wll ncrease as antennas move away from the track. dstance and the postonng results s absolutely stable wth the localzaton error of about 0.35cm. From above experments, the trackng accuracy s related to dfferent deployments of antennas. In the followng, we use an example to further dscuss such a dfference. Four RFID antennas are deployed on both sdes of the crcular track, where the coordnates of antennas 1-4 are (0,-600), (100,-600), (100,600) and (0,600) respectvely. The ntal pont of the tracked tag s (50, 0). Fg. 18 shows the result of the poston correlaton for each grd after optmzaton. The maxmum of the poston correlaton on the grd W=100,L=65 s Num100, , so the correspondng coordnate s (50, 32.5) and devates from the ground truth by 32.25cm. The error manly comes from the y-axs. Also, there exst many possble canddates because the values of Num on these grds are W,L very close to each other. However, f we deploy antennas around the montorng area, where the coordnates of antennas 1-4 are (-600,0), (0,-600), (600,0) and (0,600) respectvely, can the error from x-axs and y-axs be cancelled out each other? The maxmum on the grd W=100,L=1 s Num100, n Fg. 19, so the correspondng coordnate s (50,0.5) and devates from the ground truth by 0.5cm. Therefore, we can obtan a hgh accuracy when antennas are deployed around the survellance regon. From the experment, when we deploy antennas on the both sdes of the track, the dstance between antennas and the track should not be far, and t s reasonable to keep t wthn the 3m, otherwse there wll be a large postonng error. The man reason of declne n accuracy of localzaton s resulted from dluton of precson (DOP) [31]. Lke GPS, the accuracy of RFID trackng system depends on the geometry of reader antennas. The RFID antennas 1 and 3 n Fg. 19 le n a drecton orthogonal to that of antennas 2 and 4, gvng a relatvely small regon n whch the object tag must le wth some degree of certanty. And the object tag s X and Y coordnates are determned wth equal precson. For example, n Fg. 20 the Fg. 18. Poston ambguty whle antennas are deployed on both sdes of the montorng area. Fg. 19. The elmnaton of poston ambguty whle antennas are deployed around the montorng area. Fg. 20. Low dluton of precson. The RFID antenna 1 les n a drecton orthogonal to that of antenna 2, gvng a relatvely small regon (the ntersecton of the crcular lnes), so the poston uncertanty of the object RFID tag s small. As a result, t s a feasble way to reduce the uncertanty regon by ncreasng RFID antennas deployed n the survellance regon. Further, the antennas should be deployed around the track [32, 33] to obtan more stable and hgher postonng results under the permsson condton. C. Evaluaton for unknown track We use the scenaro of crcular track descrbed before to evaluate our method to track the unknown trajectory. The toy tran attached on a tag moves antclockwse along the track at a constant speed of 20cm/s and the ntal poston of tag s (50,0). The tag on the toy tran moves 10 laps antclockwse wthn about 157 seconds. We collect phase data durng the tme to estmate the tag s movng speed and trackng accuracy.

12 12 1) The estmated speed At frst, we evaluate the speed of tag movement. In Fg. 21, the arrow represents the velocty vector calculated n every read round. In the experment, the estmated speed has a mean of 19.11cm/s whle the ground truth speed s 20cm/s. Fg. 22 llustrates that the estmated speed tends towards stablty after more than about 200 reads are collected by four antennas. tolerance to multpath effect, whch s employed to acqure all of canddates n the survellance area. On ths bass, we calculate the double-dfference values correspondng to these canddates and select the mnmum value as the fnal poston. In addton, snce the tme between successve nventores of the same tag s short enough and an RFID tag are not movng at Fg. 21. Estmated speed for unknown track. The estmated speed has a mean of cm/s whle the ground truth speed s 20cm/s. Fg. 22. Stablty for unknown track. The estmated speed tends towards stablty after more than 200 reads are collected by four antennas. 2) The accuracy between the fxed trajectory and the ground truth In Fg. 23, each pont represents the estmated tag poston based on the collected phase n every round whle the red lne represents the ground truth. In the experment, we extract the coordnate ponts whose x-axs coordnates are wthn [-50,50], to calculate correspondng y-axs coordnates n the crcular track, whch s regarded as the ground truth n y-axs dmenson. So TrackT has a mean error dstance of 0.37cm wth the standard devaton of 0.53cm n y-axs dmenson. Smlarly, we can also acqure a mean error dstance of 0.41cm wth the standard devaton of 0.62cm n x-axs dmenson by extractng the coordnate ponts whose y-axs coordnates are wthn [-50,50]. Fnally, the mean combned error dstance s approxmately 0.55cm wth the standard devaton of 0.75cm. VI. Concluson In ths paper we manly acheve to track moble tags based on phase collected by the COTS RFID reader. The dfference of phase perodcty deduced by reported phase has strong Fg. 23. Tracked postons. Each black pont represents the localzaton result of TrackT n every read round. a hgh speed, we use a frst-order Taylor seres approxmaton to compute a tag s dsplacements every round of antenna scheduled. Moreover, the rule of reducng computaton s also proposed n ths paper to make TrackT meet real-tme performance. The results of our experment demonstrate that TrackT can be appled to track known and unknown movement trajectory wth hgh accuracy n real tme. The lmtatons and practcal deployment ssues are as follows: () Tag populaton sze. Our current desgn s to track a small amount of RFID tags at the same tme because the tme between successve nventores of the same tag wll depend on tag populaton sze. Whle large-scale RFID tags exst n the survellance zone, the movng dstance wll be large than /4 durng the tme nterval, leadng to the falure of our localzaton system. () Calbraton of antenna coordnates. Before trackng RFID tag movement trajectory, the system needs to calbrate accurate postons of each RFID antenna at frst. Ths s because the phase s very senstve to the dstance between the tag and antenna. We employ a RFID tag fxed on the known poston and a laser rangefnder wth hgh measurement accuracy to calbrate the locaton of RFID antenna, thereby mprovng localzaton accuracy. In the future, we wll rely on captured RFID tags movement trajectory to recognze human moton gestures, whch can be used n vrtual realty. REFERENCES [1] X. Yang, Y. Zhang and X. Lang, Prmate-nspred communcaton methods for moble and statc sensors and RFID tags, ACM Transactons on Autonomous and Adaptve Systems (TAAS), vol. 6, no. 4, Artcle No. 26, Oct

13 13 [2] X. Yang, S. Yu, K. Wu, Q. N, C. Janecek and J. Nordstad, Rado frequency dentfcaton: technologes, applcatons, and research ssues, Wreless Communcatons and Moble Computng, vol. 7, no. 4, pp , Jul [3] C. Floerkemeer and M. Lampe, Issues wth RFID usage n ubqutous computng applcatons, n Proc. Pervasve, 2004, pp [4] J. Brusey, C. Floerkemeer, M. Harrson and M. Fletcher, Reasonng about uncertanty n locaton dentfcaton wth RFID, n Proc. IJCAI, 2003, pp [5] L. Yang, Y. Q, J. Fang, X. Dng, T. Lu and M. L, Reasonng about uncertanty n locaton dentfcaton wth RFID, n Proc. IEEE INFOCOM, 2014, pp [6] P. Vorst and A. Zell, Fully autonomous trajectory estmaton wth long-range passve RFID, n Proc. IEEE ICRA, 2010, pp [7] L.M. N, Y. Lu, Y.C. Lau and A.P. Patl, LANDMARC: ndoor locaton sensng usng actve RFID, Wreless networks, vol. 10, no. 6, pp , Nov [8] C. Wang, H. Wu and N.F. Tzeng, RFID-based 3-D postonng schemes, n Proc. IEEE INFOCOM, 2007, pp [9] A. Bekkal, H. Sanson and M. Matsumoto, RFID ndoor postonng based on probablstc RFID map and Kalman flterng, n Proc. IEEE WMOB, 2007, pp [10] L. Shangguan, Z. L, Z. Yang, M. L and Y. Lu, Otrack: Order trackng for luggage n moble RFID systems, n Proc. IEEE INFOCOM, 2013, pp [11] J.D. Grffn and G.D. Durgn, Complete lnk budgets for backscatter-rado and RFID systems, IEEE Antennas and Propagaton Magazne, vol. 51, no. 2, pp , Apr [12] ImpnJ (2013, Sep.). Speedway revoluton reader applcaton note: Low level user data support. ImpnJ Inc. Seattle, USA. [Onlne]. Avalable: e-low-level-user-data-support [13] Y. Zhao, Y. Lu and L.M. N, Vre: Actve rfd-based localzaton usng vrtual reference elmnaton, n Proc. IEEE ICPP, 2007, pp.56. [14] C. Hekman-Wllams, B. Grant, X. Lu, Z. Zhang and P. Kumar, Accurate localzaton of RFID tags usng phase dfference, n Proc. IEEE RFID, 2010, pp [15] S. Azzouz, M. Cremer, U. Dettmar, R. Kronberger and T. Kne, New measurement results for the localzaton of UHF RFID transponders usng an Angle of Arrval (AoA) approach, n Proc. IEEE RFID, 2011, pp [16] P.V. Nktn, R. Martnez, S. Ramamurthy, H. Leland,G. Spess and K. V. S. Rao, Phase based spatal dentfcaton of UHF RFID tags, n Proc. IEEE RFID, 2010, pp [17] T. Lu, L. Yang, Q. Ln, Y. Guo and Y. Lu, Anchor-free backscatter postonng for RFID tags wth hgh accuracy, n Proc. IEEE INFOCOM, 2014, pp [18] J. Wang and D. Katab, Dude, where's my card?: RFID postonng that works wth multpath and non-lne of sght, n Proc. ACM SIGCOMM, 2013, pp [19] L. Yang, Y. Chen, X. L, C. Xao, M. L and Y. Lu, Tagoram: Real-tme trackng of moble RFID tags to hgh precson usng COTS devces, n Proc. ACM MobCom, 2014, pp [20] D.M. Dobkn, The RF n RFID: UHF RFID n Practce, Lncoln, USA: Newnes, 2012, pp [21] ImpnJ (2015, Jun.). Octane LLRP Gude verson ImpnJ Inc. Seattle, USA. [Onlne]. Avalable: ntaton [22] P. Zhang, J. Gummeson and D. Ganesan, Blnk: A hgh throughput lnk layer for backscatter communcaton, n Proc. ACM MobSys, 2012, pp [23] C. Wang, L. Xe and W. Wang, Movng tag detecton va physcal layer analyss for large-scale RFID systems, n Proc. IEEE INFOCOM, 2016, pp [24] H. Mustafa, X. Zhu, Q. L and G. Chen, Effcent medan estmaton for large scale sensor RFID systems, Internatonal Journal of Sensor Networks, vol. 12, no. 3, pp , Nov [25] G. Ferrar, F. Cappellett and R. Rahel, A smple performance analyss of RFID networks wth bnary tree collson arbtraton, Internatonal Journal of Sensor Networks, vol. 4,no.3, pp , [26] B. Hofmann-Wellenhof, H. Lchtenegger and J. Collns, Global postonng system: theory and practce, New York, USA: Sprnger Scence & Busness Meda, 2013, pp [27] A.F.C. Errngton, B.L.F. Daku and A.F. Prugger, Intal poston estmaton usng RFID tags: A least-squares approach, IEEE Trans. Instrumentaton and Measurement, vol. 59, no. 11, pp , Nov [28] B. Gu, V. S. Sheng, K. Y. Tay, W. Romano, and S. L Incremental support vector learnng for ordnal regresson. Wreless networks, vol. 26, no. 7, pp , July [29] S. Sen, J. Lee, K. H. Km and P. Congdon, Avodng multpath to revve nbuldng WF localzaton, n Proc. ACM MobSys, 2013, pp [30] J. Xong and K. Jameson, ArrayTrack: A Fne-Graned Indoor Locaton System, n Proc. NSDI, 2013, pp [31] R. B. Langley, Dluton of precson, GPS world, vol. 10, no. 5, pp , May [32] Z. Wang, N. Ye, R. Malekan, R. Wang, Peng L,"TMcroscope: Behavor Percepton Based on the Slghtest RFID Tag Moton", Elektronka r Elektrotechnka, Vol.22, No.2, pp ,2016. [33] N. Ye, Z. Q. Wang, R. Malekan, Q. Ln, R. Wang, "A Method for Drvng Route Predctons Based on Hdden Markov Model", Mathematcal Problems n Engneerng, Vol.2015, pp.1-12, Zhongqn Wang (STM 16) receved the Master degree n Computer Software and Theory from Nanjng Unversty of Posts and Telecommuncatons, Nanjng, Chna n He s currently pursung the Ph.D. degree at Nanjng Unversty of Posts and Telecommuncatons. In 2016, he became a Student Member of IEEE. Hs research nterest ncludes Rado Frequency Identfcaton (RFID), sensor network and fancy Human Interacton. Nng Ye receved the B.S. degree n Computer Scence from Nanjng Unversty n 1994, the M.S. degree n School of Computer & Engneerng from Southeast Unversty n 2004, and the Ph.D. degree n Insttute of Computer Scence from Nanjng Unversty of Posts and Telecommuncatons n She s currently a Professor there. In 2010, Nng Ye worked as a Vstng Scholar and Research Assstant n the Department of Computer Scence, Unversty of Vctora, Canada. Her research nterests nclude nformaton processng n wreless networks and Internet of Thngs. She s a senor member of Chnese Computer Federaton (CCF). Reza Malekan receved the B.Eng. degree n computer engneerng, the M.Eng. degree n telecommuncatons engneerng, and the Ph.D. degree n computer scence. He s currently a Senor Lecturer wth the Department of Electrcal, Electronc and Computer Engneerng, Unversty of Pretora, South Afrca. Hs research les n the area of advanced sensor networks, Internet of Thngs, and moble communcatons. He s also a Professonal Member of the Brtsh Computer Socety and Chartered Engneer. Fu Xao (M 12) receved the Ph.D Degree n Computer Scence and Technology from Nanjng Unversty of Scence and Technology, Nanjng, Chna, n He s