UNATTENDED ground sensors (UGS) are widely used in

Transcription

1 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE Trget Detetion n Clssifition Using Seismi n PIR Sensors Xin Jin, Stuent Member, IEEE, Soumly Srkr, Asok Ry, Fellow, IEEE, Shlbh Gupt, Member, IEEE, n Thygrju Dmrl, Senior Member, IEEE Abstrt Unttene groun sensors (UGS) re wiely use to monitor humn tivities, suh s peestrin motion n etetion of intruers in seure region. Effiy of UGS systems is often limite by high flse lrm rtes, possibly ue to inequies of the unerlying lgorithms n limittions of onbor omputtion. In this regr, this pper presents wvelet-bse metho for trget etetion n lssifition. The propose metho hs been vlite on t sets of seismi n pssive infrre sensors for trget etetion n lssifition, s well s for pylo n movement type ientifition of the trgets. The propose metho hs the vntges of fst exeution time n low memory requirements n is potentilly well-suite for rel-time implementtion with onbor UGS systems. Inex Terms Feture extrtion, pssive infrre sensor, seismi sensor, symboli ynmi filtering, trget etetion n lssifition. I. INTRODUCTION UNATTENDED groun sensors (UGS) re wiely use in inustril monitoring n militry opertions. Suh UGS systems re usully lightweight evies tht utomtilly monitor the lol tivities in-situ, n trnsfer trget etetion n lssifition reports to the proessing enter t higher level of hierrhy. Commerilly vilble UGS systems mke use of multiple sensing molities (e.g., ousti, seismi, pssive infrre, mgneti, eletrostti, n vieo). Effiy of UGS systems is often limite by high flse lrm rtes beuse the onbor t proessing lgorithms my not be ble to orretly isriminte ifferent types of trgets (e.g., humns from nimls) [1]. Power onsumption is ritil onsiertion in UGS systems. Therefore, power-effiient Mnusript reeive September 14, 211; revise Otober 2, 211; epte November 15, 211. Dte of publition November 23, 211; te of urrent version April 2, 212. This work ws supporte in prt by the U.S. Army Reserh Lbortory n the U.S. Army Reserh Offie uner Grnt W911NF Any opinions, finings n onlusions or reommentions expresse in this publition re those of the uthors n o not neessrily reflet the views of the sponsoring genies. The ssoite eitor oorinting the review of this pper n pproving it for publition ws Prof. Rlph Etienne-Cummings. X. Jin, S. Srkr, n A. Ry re with the Deprtment of Mehnil Engineering, Pennsylvni Stte University, University Prk, PA 1682 USA (e-mil: xinjin@psu.eu; svs5464@psu.eu; xr2@psu.eu). S. Gupt is with the Deprtment of Eletril n Computer Engineering, University of Connetiut, Storrs, CT 6269 USA (e-mil: shlbh.gupt@engr.uonn.eu). T. Dmrl is with the U.S. Army Reserh Lbortory, Aelphi, MD 2783 USA (e-mil: thygrju.mrl@us.rmy.mil). Color versions of one or more of the figures in this pper re vilble online t Digitl Objet Ientifier 1.119/JSEN X/$ IEEE sensing molities, low-power signl proessing lgorithms, n effiient methos for exhnging informtion between the UGS noes re neee [2]. In the etetion n lssifition problem t hn, the trgets usully inlue humn, vehiles n nimls. For exmple, isriminting humn footstep signls from other trgets n noise soures is hllenging tsk, beuse the signl-to-noise rtio (SNR) of footsteps ereses rpily with the istne between the sensor n the peestrin. Furthermore, the footstep signls my vry signifintly for ifferent people n environments. Often the wek n noiseontminte signtures of humns n light vehiles my not be lerly istinguishble from eh other, in ontrst to hevy vehiles tht rite lou signtures [3], [4]. Seismi sensors re wiely use for personnel etetion, beuse they re reltively less sensitive to Doppler effets n environment vritions, s ompre to ousti sensors [5]. Current personnel etetion methos, bse on seismi signls, re lssifie into three tegories, nmely, time omin [6], frequeny omin [3], [4], [7], n time-frequeny omin [5], [8] [1]. Generlly, time-omin nlysis my not be ble to etet trgets very urtely beuse of the interfering noise, omplite signl wveforms, n vritions of the terrin [5]. On the other hn, ury of frequeny omin methos my be egre ue to unerlying non-sttionrity in the observe signl. Therefore, reent reserh hs relie on time-frequeny omin (e.g. wvelet trnsform-bse) methos beuse of their enoising n loliztion properties. Pssive Infrre (PIR) sensors hve been wiely use in motion etetors, where the PIR signls re usully quntize into two sttes, i.e., on n off. PIR signls ontin isrimintive informtion in the time-frequeny omin n re well-suite for UGS systems ue to low power onsumption. Although PIR sensors hve been use for etetion n loliztion of moving trgets [11], similr efforts for trget lssifition hve not been pprently reporte in open literture. The work reporte in this pper mkes use of wveletbse feture extrtion metho, lle Symboli Dynmi Filtering (SDF) [12] [14]. The SDF-bse feture extrtion lgorithm mitigtes the noise by using wvelet nlysis, ptures the essentil signtures of the originl signls in the time-frequeny omin, n genertes robust low-imensionl feture vetors for pttern lssifition [15]. This pper resses the problem of trget etetion n lssifition using seismi n PIR sensors tht monitor the infiltrtion of humns, light vehiles n omesti nimls for borer

2 Report Doumenttion Pge Form Approve OMB No Publi reporting buren for the olletion of informtion is estimte to verge 1 hour per response, inluing the time for reviewing instrutions, serhing existing t soures, gthering n mintining the t neee, n ompleting n reviewing the olletion of informtion. Sen omments regring this buren estimte or ny other spet of this olletion of informtion, inluing suggestions for reuing this buren, to Wshington Hequrters Servies, Diretorte for Informtion Opertions n Reports, 1215 Jefferson Dvis Highwy, Suite 124, Arlington VA Responents shoul be wre tht notwithstning ny other provision of lw, no person shll be subjet to penlty for filing to omply with olletion of informtion if it oes not isply urrently vli OMB ontrol number. 1. REPORT DATE 2 APR REPORT TYPE 3. DATES COVERED to TITLE AND SUBTITLE Trget Detetion n Clssifition Using Seismi n PIR Sensors 5. CONTRACT NUMBER 5b. GRANT NUMBER 5. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) Pennsylvni Stte University,Deprtment of Mehnil Engineering,University Prk,PA, PERFORMING ORGANIZATION REPORT NUMBER 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 1. SPONSOR/MONITOR S ACRONYM(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approve for publi relese; istribution unlimite 13. SUPPLEMENTARY NOTES 11. SPONSOR/MONITOR S REPORT NUMBER(S) 14. ABSTRACT Unttene groun sensors (UGS) re wiely use to monitor humn tivities, suh s peestrin motion n etetion of intruers in seure region. Effiy of UGS systems is often limite by high flse lrm rtes, possibly ue to inequies of the unerlying lgorithms n limittions of onbor omputtion. In this regr, this pper presents wvelet-bse metho for trget etetion n lssifition. The propose metho hs been vlite on t sets of seismi n pssive infrre sensors for trget etetion n lssifition s well s for pylo n movement type ientifition of the trgets. The propose metho hs the vntges of fst exeution time n low memory requirements n is potentilly well-suite for rel-time implementtion with onbor UGS systems. 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF ABSTRACT. REPORT unlssifie b. ABSTRACT unlssifie. THIS PAGE unlssifie Sme s Report (SAR) 18. NUMBER OF PAGES NAME OF RESPONSIBLE PERSON Stnr Form 298 (Rev. 8-98) Presribe by ANSI St Z39-18

3 171 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 212 Fig. 1. Illustrtion of the test senrio with three sensor sites. seurity. The mjor ontributions of the pper re s follows: 1) Formultion of hierrhil struture for trget etetion n lssifition. 2) Experimentl vlition of the SDF-bse feture extrtion metho on seismi n PIR sensor t. 3) Performne evlution of using seismi n PIR sensors in trget pylo n movement type ientifition. The pper is orgnize into five setions inluing the present one. Setion II esribes n formultes the problem of trget etetion n lssifition. Setion III presents the proeure of feture extrtion from sensor time-series. Setion IV esribes the etils of the propose metho n the results of fiel t nlysis. The pper is onlue in Setion V long with reommentions for future reserh. II. PROBLEM DESCRIPTION AND FORMULATION The objetive is to etet n lssify ifferent trgets (e.g., humns, vehiles, n nimls le by humn), where seismi n PIR sensors re use to pture the hrteristi signtures. For exmple, in the movement of humn or n niml ross the groun, osilltory motions of the boy ppenges provie the respetive hrteristi signtures. The seismi n PIR sensor t, use in this nlysis, were ollete on multiple ys from test fiels on wsh (i.e., the ry be of n intermittent reek) n t hoke point (i.e., ple where the trgets re fore to go ue to terrin iffiulties). During multiple fiel tests, sensor t were ollete for severl senrios tht onsiste of trgets wlking long n pproximtely 15 meters long tril, n returning long the sme tril to the strting point. Figure 1 illustrtes typil t olletion senrio. The trgets onsiste of humns (e.g., mle n femle), nimls (e.g., onkeys, mules, n horses), n ll-terrin vehiles (ATVs). The humns wlke lone n in groups with n without bkpks; the nimls were le by their humn hnlers (simply enote s niml in the sequel) n they me runs with n without pylos; n ATVs move t ifferent spees (e.g., 5 mph n 1 mph). Exmples of the test senrios with ifferent trgets re shown in Fig. 2. There were three sensor sites, eh equippe with seismi n PIR sensors. The seismi sensors (geophones) were burie pproximtely 15 m eep unerneth the soil surfe, n the PIR sensors were ollote with the respetive seismi sensors. All trgets psse by the sensor sites t istne of pproximtely 5 m. Signls from both sensors were quire t smpling frequeny of 1 khz. () (b) () Fig. 2. Exmples of test senrios with ifferent trgets. (). (b) Vehile. () le by humn. Detetion Clssifition Fig. 3. Seismi signls Vehile Trget present Wlking/Running? Crrying pylo? Feture extrtion / Trget bsent Crrying pylo? Tree struture formultion of the etetion & lssifition problem. The tree struture in Fig. 3 shows how the etetion n lssifition problem is formulte. In the etetion stge, the pttern lssifier etets the presene of moving trget ginst the null hypothesis of no trget present; in the lssifition stge, the pttern lssifiers isriminte mong ifferent trgets, n subsequently ientify the movement type n/or pylo of the trgets. While the etetion system shoul be robust to reue the flse lrm rtes, the lssifition system must be suffiiently sensitive to isriminte mong ifferent types of trgets with high fielity. In this ontext, feture extrtion plys n importnt role in trget etetion n lssifition beuse the performne of lssifiers lrgely epens on the qulity of the extrte fetures. In the lssifition stge, there re multiple lsses (i.e., humns, nimls, n vehiles); n the signture of the vehiles is istint from those of the other two lsses. Therefore, this problem is formulte into two-lyer lssifition proeure. A binry lssifition is performe to etet the presene of trget n then to ientify whether the trget is vehile or humn/niml. Upon reognizing the trget s humn/niml, nother binry lssifition is performe to etermine its speifi lss. More informtion oul be erive upon reognition of the trget type. For exmple, if the trget

4 JIN et l.: TARGET DETECTION AND CLASSIFICATION USING SEISMIC AND PIR SENSORS 1711 SDF-bse feture extrtion Time shifts () 1 Wvelet trnsform b 1 Sles 5 Prtition surfe Wvelet surfe 1 Shifts (b) Symboliztion Sles 6 b b 5 b b b b b Shifts () Conversion to stte imge b Pttern genertion PFSA onstrution O2 O Stte inex (f) (e) () Fig. 4. Overview of the SDF-bse feture extrtion lgorithm. () Sensor time series t. (b) Prtition of the wvelet oeffiients. () Symbolize wvelet imge ( setion). () Feture extrtion from stte imge. (e) 4-stte PFSA. (f) Stte probbility vetor. is reognize s humn, then further binry lssifitions re performe to ientify if the humn is running or wlking, n if the humn is rrying pylo or not. III. SYMBOLIC DYNAMICS-BASED FEATURE EXTRACTION A key step in trget etetion n lssifition is feture extrtion from sensor signls, whih is omplishe by symboli ynmi filtering (SDF) in this pper. While the etils of SDF hve been reporte in erlier publitions [12] [15], this setion briefly reviews the unerlying onepts of feture extrtion from sensor time series for ompleteness of this pper. A. Trnsformtion of Time Series to Wvelet Domin A ruil step in SDF is prtitioning of the trnsforme t spe for symbol sequene genertion. In wvelet-bse prtitioning, the time series is first trnsforme s set of wvelet oeffiients t ifferent time shifts n sles, where the hoie of the wvelet bsis funtion epens on the timefrequeny hrteristis of the unerlying signl, n the (finitely mny) wvelet sles re lulte s follows: α i = F f p i t (1) where F is the enter frequeny [16] tht hs the mximum moulus in the Fourier trnsform of the signl; n f p i s re obtine by hoosing the lolly ominnt frequenies in the Fourier trnsform. Figure 4 shows n illustrtive exmple of trnsformtion of the time series (Fig. 4()) to (two-imensionl) wvelet imge (Fig. 4(b)). The mplitues of the wvelet oeffiients over the sle-shift omin re plotte s surfe. Subsequently, symboliztion of this wvelet surfe les to the formtion of symboli imge s shown in Fig. 4(). B. Symboliztion of Wvelet Surfe Profiles This setion presents prtitioning of the wvelet surfe profile, s shown in Fig. 4(b), whih is generte by the oeffiients over the two-imensionl sle-shift omin, for onstrution of the symboli imge in Fig. 4(). The x y oorintes of the wvelet surfe profiles enote the shifts n the sles respetively, n the z-oorinte enotes the surfe height s pixel vlues of the wvelet oeffiients. The wvelet surfe profiles re prtitione suh tht the orintes between the mximum n minimum of the oeffiients long the z-xis re ivie into regions by ifferent plnes prllel to the x y plne. For exmple, if the lphbet is hosen s = {, b,, }, i.e., = 4, then three prtitioning plnes ivie the orinte (i.e., z-xis) of the surfe profile into four mutully exlusive n exhustive regions, s shown in Figure 4(b). These isjoint regions form prtition, where eh region is lbele with one symbol from the lphbet. If the intensity of pixel is lote in prtiulr region, then it is oe with the symbol ssoite with tht region. As suh, symbol from the lphbet is ssigne to eh pixel orresponing to the region where its intensity flls. Thus, the two-imensionl rry of symbols,

5 1712 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 212 lle symbol imge, is generte from the wvelet surfe profile, s shown in Figure 4(). The surfe profiles n be prtitione by using ifferent prtitioning methos. If the prtitioning plnes re seprte by equl-size intervls, then it is lle the uniform prtitioning (UP). However, the prtitioning woul be more resonble if the informtion-rih regions of t set re prtitione finer n those with sprse informtion re prtitione orser. To hieve this objetive, the mximum entropy prtitioning (MEP) [13], [14] hs been opte suh tht the entropy of the generte symbols is mximize. The proeure for seletion of the lphbet size, followe by genertion of MEP, hs been reporte in [13], [14]. In generl, the hoie of lphbet size epens on speifi t set. The prtitioning of wvelet surfe profiles to generte symboli representtions enbles robust feture extrtion, n symboliztion lso signifintly reues the memory requirements. For the purpose of pttern lssifition, the referene t set is prtitione with lphbet size n is subsequently kept onstnt. In other wors, the struture of the prtition is fixe t the referene onition n this prtition serves s the referene frme for subsequent t nlysis [12]. C. Conversion of the Symbol Imge to the Stte Imge This setion presents onstrution of probbilisti finite stte utomton (PFSA) for feture extrtion bse on the symbol imge generte from wvelet surfe profile. For nlysis of (one-imensionl) time series, the sttes of PFSA represent ifferent ombintions of bloks of symbols on the symbol sequene n the eges represent the trnsition probbilities between these bloks [12]. Therefore, for nlysis of (one imensionl) time series, the sttes enote ll possible symbol bloks (i.e., wors) within winow of ertin length. The notion of sttes is now extene for nlysis of wvelet surfe profiles vi onstrution of stte imge from symbol imge. In generl, the omputtionl requirements inrese with the number Q of sttes, whih must be onstrine for rel-time pplitions. As Q inreses with the winow size W n the lphbet size, probbilisti stte ompression metho is employe, whih hooses m most probble symbols from eh stte s representtion of tht prtiulr stte. Stte ompression must preserve suffiient informtion s neee for pttern lssifition, lbeit possibly lossy oing of the wvelet surfe profile. In this metho, eh stte onsisting of l l symbols is ompresse to wor of length m < l 2 symbols by hoosing the top m symbols tht hve the highest probbility of ourrene. (Note: If two symbols hve the sme probbility of ourrene, then either symbol my be preferre with equl probbility.) This proeure reues the stte set Q to n effetive smller set O {o 1, o 2,...,o O } tht enbles mpping of two or more ifferent onfigurtions in winow W to single stte. For exmple, if =4, W =4n m = 2, then the stte ompression reues the totl number of sttes to O m = 16 inste of 256. The hoie of, l n m epens on pplition omins, noise level, n the vilble omputtionl power, n is me by n pproprite treoff between robustness to noise n pbility to etet smll hnges. For exmple, lrge lphbet my be noise-sensitive while smll lphbet my miss the informtion of signl ynmis. This issue is isusse further in Subsetion IV-A.2. D. Constrution of PFSA n Pttern Genertion A probbilisti finite stte utomton (PFSA) is onstrute suh tht the sttes of the PFSA re elements of the ompresse stte set O n the eges re trnsition probbilities between these sttes. The trnsition probbilities re efine s: N(o l, o k ) (o k o l ) = k =1,2,..., O N(o o l, o k O (2) l, o k ) where N(o l, o k ) is the totl ount of events when o k ours jent to o l in the iretion of motion. The lultion of these trnsition probbilities follows the priniple of sliing blok oe [17]. A trnsition from the stte o l to the stte o k ours if o k lies jent to o l in the positive iretion of motion. Subsequently, the ounter moves to the right n to the bottom (row-wise) to over the entire stte imge, n the trnsition probbilities (o k o l ) o l, o k O re ompute using Eq. (2). Therefore, for every stte on the stte imge, ll stte-to-stte trnsitions re ounte, s shown in Fig. 4(). For exmple, the otte box in the bottom-right orner ontins three jent pirs, implying the trnsitions o 1 o 2, o 1 o 3,no 1 o 4 n the orresponing ounter of ourrenes N(o 1, o 2 ), N(o 1, o 3 ),nn(o 1, o 4 ), respetively, re inrese by one. This proeure genertes the stte-trnsition probbility mtrix of the PFSA given s: (o 1 o 1 )... (o O o 1 ) = (3) (o 1 o O )... (o O o O ) where [π jk ] with π jk = (o k o j ).Note:π jk j, k {1, 2,... O } n k π jk = 1 j {1, 2,... O }. In orer to extrt low-imensionl feture vetor, the sttionry stte probbility vetor p is obtine s the left eigenvetor orresponing to the unity eigenvlue of the stohsti trnsition mtrix. The stte probbility vetors p serve s the feture vetors n re generte from ifferent t sets from the orresponing stte trnsition mtries. These feture vetors re enote s ptterns inthis pper. IV. RESULTS OF FIELD DATA ANALYSIS Fiel t were ollete in the senrio illustrte in Fig. 1. Multiple experiments were me to ollet t sets of ll three lsses, i.e., humn, vehile n niml. The t were ollete over three ys t ifferent sites. A brief summry is given in Tble I showing the number of runs of eh lss. Eh t set, quire t smpling frequeny of 1 khz, hs t points tht orrespon to 1 seons of the experimenttion time. In orer to test the pbility of the propose lgorithm for trget etetion, nother t set

6 JIN et l.: TARGET DETECTION AND CLASSIFICATION USING SEISMIC AND PIR SENSORS 1713 TABLE I NUMBER OF FEATURE VECTORS FOR EACH TARGET CLASS Dy 1 Dy 2 Dy 3 Totl No trget Vehile ws ollete with no trget present. The problem of trget etetion is then formulte s binry pttern lssifition, where no trget present orrespons to one lss, n trget present (i.e., humn, vehile or niml) orrespons to the other lss. The t sets, ollete by the hnnel of seismi sensors tht re orthogonl to the groun surfe n the PIR sensors tht re ollote with the seismi sensors, re use for trget etetion n lssifition. For omputtionl effiieny, the t were ownsmple by ftor of 1 with no pprent loss of informtion. Figure 5 epits the flow hrt of the propose etetion n lssifition lgorithm tht is onstrute bse on the theories of symboli ynmi filtering (SDF) n support vetor mhines (SVM) [18]. The propose lgorithm onsists of four min steps, nmely, signl preproessing, feture extrtion, etetion, n lssifition, s shown in Fig. 5. In the signl onitioning step, the DC omponent (i.e., the onstnt offset) of seismi signl is eliminte by subtrting the verge vlue n the resulting (zero-men) signl is normlize to unit-vrine with ivision by its stnr evition. The rtionle for normliztion to unit vrine is to mke pttern lssifition inepenent of the signl mplitue n ny isrimintion shoul be solely texture-epenent. For exmple, the mplitue of the seismi signl of n niml with hevy pylo wlking fr wy oul be omprble to tht of peestrin pssing by t loser istne, lthough these two signls re of ifferent texture. However, for PIR signls, only the DC omponent is remove n the normliztion is not rrie out beuse the rnge of the PIR signls is not hnge uring the fiel test experiments. In the feture extrtion step, SDF ptures the signtures of the preproesse sensor time-series for representtion s lowimensionl feture vetors. Bse on the spetrl nlysis of the ensemble of seismi t t hn, series of pseuofrequenies from the 1-2 Hz bns hve been hosen to generte the sles for wvelet trnsform, beuse these bns ontin very lrge prt of the footstep energy [8]. Similrly, series of pseuo-frequenies from the.2-2. Hz bns hve been hosen for PIR signls to generte the sles. Upon genertion of the sles, ontinuous wvelet trnsforms (CWT) re performe with n pproprite wvelet bsis funtion on the seismi n PIR signls. The wvelet bsis b7isusefor seismi signls sine it mthes the impulsive shpe of seismi signls very well, n b1 is use for the PIR se sine PIR signls re lose to squre wves. A mximum-entropy wvelet surfe prtitioning is then performe. Seletion of the lphbet size epens on the hrteristis of the signl; while smll lphbet is robust ginst noise n environmentl vritions, lrge lphbet hs more isriminnt Fig. 5. Signl preproessing Feture extrtion Detetion Clssifition (Type) Clssifition (Detils) Sensor signl Remove DC omponent Normlize to unit vrine Time series Symboli ynmi filtering (SDF) Ptterns Pttern lssifier # 1 Trget present Pttern lssifier # 2 / Pttern lssifier # 3 Pttern lssifier Wlking? Running? Crrying Pylo? Pttern lssifier Crrying Pylo? No trget Vehile Flow hrt of the problem of trget etetion n lssifition. power for ientifying ifferent objets. The sme lphbet is use for both trget etetion n lssifition. The issues of optimiztion of the lphbet size n t set prtitioning re not resse in this pper. Subsequently, the extrte low-imensionl ptterns re use for trget etetion n lssifition. One pttern is generte from eh experiment, n the trining ptterns re use to generte the seprting hyperplne in SVM. A. Performne Assessment Using Seismi Dt This setion presents the lssifition results using the ptterns extrte from seismi signls using SDF. The leveone-out ross-vlition metho [18] hs been use in the performnessessment of seismi t. Sine the seismi sensors re not site-inepenent, they require prtil informtion of the test site, whih is obtine from the trining set in the ross-vlition. Results of trget etetion n lssifition, movement type n trget pylo ientifition re reporte in this setion. 1) Trget Detetion n Clssifition: Figure 6 shows the normlize seismi sensor signls (top row) n the orresponing feture vetors (bottom row) extrte by SDF of the three lsses of trgets n the no trget se. In the top row of Fig. 6, the unit of the orinte xes is imensionless ue to normliztion of the seismi signls, where the originl t were reore in the unit of volt by mirophones for storge in igitize formt. Eh feture vetor in the bottom row of Fig. 6 onsists of 8 elements sine the lphbet size =8

7 1714 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 212 Normlize seismi signl Stte inex () Normlize seismi signl Stte inex (b) Normlize seismi signl Stte inex () Normlize seismi signl Stte inex () Fig. 6. () No trget. (b) Vehile. (). (). Exmples of seismi sensor mesurements (top) n the orresponing feture vetors extrte by SDF of the four lsses (bottom). TABLE II CONFUSION MATRICES OF THE LEAVE-ONE-OUT CROSS-VALIDATION RESULTS USING SDF AND KURTOSIS ANALYSIS SDF No trget Vehile No trget Vehile Kurtosis No trget Vehile No trget Vehile n the sum of ll the elements in eh feture vetor is 1. It is observe tht the feture vetors re quite ifferent mong no trget, vehile n humn/niml se. The feture vetors of humn n niml re similr n yet still istinguishble. In the feture vetor plots in Fig. 6, the sttes with smll inex number orrespons to the wvelet oeffiients with lrge vlues, n vie vers. For the purpose of omprtive evlution, kurtosis nlysis [6], benhmrking tehnique of footstep etetion, is lso use for trget etetion n lssifition. Kurtosis nlysis is useful for footstep etetion beuse the kurtosis vlue is muh higher in the presene of impulsive events (i.e., trget present) thn the se of no trget [6]. The results of SDF n kurtosis nlysis re shown in Tble II using onfusion mtries, where the rows re the tul lsses n the olumns re the preite lsses. Similr nottions re followe in the sequel in Tbles IV, V, n VI. The she re in Tble II represents the onfusion mtries of trget lssifition. The etetion n lssifition ury is summrize in Tble III. It is observe kurtosis nlysis hs slightly worse but omprble performne with SDF TABLE III COMPARISON OF THE DETECTION AND CLASSIFICATION ACCURACY BY USING SDF AND KURTOSIS ANALYSIS Clssifition Detetion versus Vehile versus Others SDF 97.% 99.1% 97.2% Kurtosis 92.4% 93.1% 55.6% TABLE IV CONFUSION MATRICES OF THE LEAVE-ONE-OUT CROSS-VALIDATION RESULTS FOR MOVEMENT TYPE IDENTIFICATION Wlking Running Wlking 47 1 Running 5 13 TABLE V CONFUSION MATRICES OF THE LEAVE-ONE-OUT CROSS-VALIDATION RESULTS FOR TARGET PAYLOAD IDENTIFICATION no pylo pylo no pylo pylo no pylo pylo no pylo pylo 2 28 in trget etetion n vehile lssifition, wheres SDF outperforms kurtosis nlysis in istinguishing humn from niml. The exeution of the MATLAB oe tkes 2.27 seons n MB of memory for SDF n SVM on esktop omputer to proess t set of points n perform pttern lssifition with the following prmeters: lphbet size =8, number of sles α =4, winow size l l = 2 2, number of most probble symbol m = 1, n qurti

8 JIN et l.: TARGET DETECTION AND CLASSIFICATION USING SEISMIC AND PIR SENSORS 1715 Normlize seismi signl TABLE VI CONFUSION MATRIX OF THE THREE-WAY CROSS-VALIDATION No trget Wlking Running No trget 11 Wlking Running Stte inex () Normlize seismi signl Stte inex (b) Fig. 7. Exmples of seismi sensor mesurements (top) n the orresponing feture vetors extrte by SDF (bottom) for humn wlking n running. () Wlking. (b) Running. kernel for SVM. Pttern lssifition onsumes bout 8% of the totl exeution time beuse by using leve-one-out ross-vlition, the pttern lssifier nee to be trine with ll the remining ptterns (e.g., 235 in etetion stge). The hoie of qurti kernel in SVM improves the performne of the lssifier; however, it lso inreses the omputtion in trining the lssifier. It is expete the exeution time n memory will be reue signifintly if fewer trining ptterns re use. 2) Movement Type Ientifition: Upon reognition of humn, more informtion n be erive by performing nother binry lssifition to ientify whether the humn is running or wlking. The physil explntions re: i) the ene (i.e., intervl between events) of humn wlking is usully lrger thn the ene of humn running; ii) the impt of running on the groun is muh stronger thn tht of wlking, n it tkes longer for the osilltion to ey. Figure 7 shows the seismi signl n orresponing feture vetors of humn wlking n running. The feture vetors of humn wlking n running re very ifferent from eh other, whih is ler inition tht the SDF-bse feture extrtion metho is ble to pture these fetures (ene n impt). It is note tht the feture vetors shown in Fig. 7 re ifferent from those in Fig. 6 beuse ifferent prtitions re use in the trget lssifition n movement type ientifition stges. Ielly, the ientifition of movement type shoul be performe bse on the results of humn lssifition. However, in orer to ssess the performne of SDF in this prtiulr pplition, binry lssifition between humn wlking n humn running is iretly performe. The results re liste in Tble IV, where the propose feture extrtion lgorithm n SVM re ble to ientify the humn movement type with n ury of bout 91%. As stte in Subsetion III-C s well s in erlier publitions [13], [14], the lphbet in the SDF lgorithm plys n importnt role for trget etetion n lssifition. An exmple illustrting the effets of the lphbet size on humn movement type ientifition is presente in Fig. 9, where the humn movement type ientifition ws performe with vrying from 2 to 2. It is seen tht tht the lssifition ury is onsistent within the rnge of from 2 to 2. The rtionle is tht the informtion loss inreses with smller n robustness to noise ereses with lrger. 3) Trget Pylo Ientifition: Similr with the movement type ientifition shown bove, the trget pylo informtion n lso be erive by performing nother binry lssifition for both niml n humn trgets. Figure 8 shows the seismi signls n feture vetors of humn/niml with n without pylo exmples. It is observe tht the feture vetors extrte by SDF hs lrge inter-lss seprtion while smll intr-lss vrine, n yet the intr-lss ifferenes between the with pylo n without pylo ses re still istinguishble. Tble V shows the results of the humn/nml pylo ientifition. The she re in Tble V represents the pylo ientifition. It is seen tht the propose metho is ble to istinguish humn from niml with high ury (97.3%). The pylo ientifition result is lso resonble (humn: 81.3%, niml: 79.1%); however, more thn hlf smples in the humn with pylo n niml without pylo ses re inorretly lssifie. Three ftors my ontribute to low lssifition rte for these two lsses: i) the pylos re not the sme throughout ll the experiments; ii) the weight of the pylo is only smll frtion of the weight of humn/niml trget, so ifferene between the two lsses (with pylo/without pylo) re not obvious; iii) Unblne number of smples in eh lss. The first two issues re relte with t olletion; the lst issue my be resolve by inresing the weight of the lss with fewer smples when generting the seprting hyperplne in SVM. B. Performne Assessment Using PIR Dt PIR sensors re wiely use for motion etetion. In most pplitions, the signls from PIR sensors re use s isrete vribles (i.e., on or off). This my work for trget etetion, but will not work well for trget lssifition beuse the time-frequeny informtion is lost in the isretiztion. In this pper, the PIR signls re onsiere to be ontinuous signls, n ontinuous wvelet trnsform (CWT) is use to revel the istintion mong ifferent types of trgets in the time-frequeny omin. Sine PIR sensor oes not emit n

9 1716 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 212 Normlize seismi signl Stte inex () Normlize seismi signl Stte inex (b) Normlize seismi signl Stte inex () Normlize seismi signl Stte inex () Fig. 8. () without pylo. (b) with pylo. () without pylo. () with pylo. Exmples of seismi sensor mesurements (top) n the orresponing feture vetors extrte by SDF (bottom) for pylo ientifition. Aury (%) Fig Alphbet size Effet of lphbet size on humn movement type ientifition. infrre bem but merely pssively epts inoming infrre rition, it is less sensitive to environmentl vritions (i.e., vrition in test sites) thn the seismi sensor tht is grounbse. A three-wy ross-vlition [18] is use for the performne ssessment of PIR t. The t re ivie into three sets by te (i.e., Dy 1, Dy 2, n Dy 3) n three ifferent sets of experiments re performe: 1) Trining: Dy 1 + Dy 2; Testing: Dy 3 2) Trining: Dy 1 + Dy 3; Testing: Dy 2 3) Trining: Dy 2 + Dy 3; Testing: Dy 1. Trining n testing on feture vetors from ifferent ys is very meningful in prtie. In eh run of the rossvlition, no prior informtion is ssume for the testing site or the testing t. The lssifiers pbility to generlize to n inepenent t set is thoroughly teste in the threewy ross-vlition. In this setion, four types of trgets re onsiere, nmely, no trget, humn wlking, humn running, n niml le by humn. Following Fig. 5, the following ses re teste: 1) Detetion of trget presene ginst trget bsene; 2) Clssifition of trget type, i.e., vs. ; 3) Clssifition of trget movement type (i.e., wlking vs. running) upon reognition of the trget s humn. Figure 1 shows the PIR sensor mesurements (top) n the orresponing feture vetors extrte by SDF (bottom) of the four lsses. For the no trget se, the PIR signl flututes roun zero n no informtion is embee in the wvelet oeffiients, thus the sttes in the mile (i.e., sttes 3-1) re oupie; wheres for the trget present ses, the PIR sensors re exite by the presene of the trgets, so sttes 1-2 n tht orrespon to the rests n troughs in the PIR signls re more populte thn other sttes. ThefollowingprmetersreuseinSDFnSVMfor proessing the PIR signls: lphbet size =12, number of sles α =3, winow size l l = 2 2, number of most probble symbol m = 1, n qurti kernel for SVM. The exeution of SDF n SVM tkes 1.13 seons n MB of memory on esktop omputer to proess t set of points, whih is ler inition of the rel-time implementtion pbility for onbor UGS systems. Tble VI shows the onfusion mtrix of the three-wy ross-vlition results using PIR sensors. The she re represents the trget lssifition stge. It is seen in Tble VI tht the propose feture extrtion lgorithm works very well with the PIR sensor; the trget etetion ury is 99.5%, the humn/niml lssifition ury is 91.7%, n the humn movement type lssifition ury is 79.3%. Leve-oneout ross-vlition usully unerestimtes the error rte in generliztion beuse more trining smples re vilble; it is expete tht the lssifition ury will further improve for the PIR signls if leve-one-out ross-vlition is use. C. Fiel Deployment of Seismi n PIR Sensors Seismi n PIR sensors hve their own vntges n isvntges for trget etetion n lssifition. The seismi sensor is omniiretionl n hs long rnge of etetion (upto7m) [1], wheres PIR sensor hs typil rnge of less thn 6 m n hs limite fiel of view (less thn 18 ), whih restrits the sensor from eteting trget moving behin it. The seismi sensor is not site-inepenent n is

10 JIN et l.: TARGET DETECTION AND CLASSIFICATION USING SEISMIC AND PIR SENSORS 1717 PIR sensor signl PIR sensor signl PIR sensor signl PIR sensor signl Stte inex Stte inex Stte inex Stte inex () (b) () () Fig. 1. Exmples of PIR sensor mesurements (top) n the orresponing feture vetors extrte by SDF (bottom) of the four lsses. () No trget. (b) wlking. () running. () le by humn. vulnerble to vritions in sensor sites, wheres PIR sensor merely pssively epts the inoming infrre rition n is inepenent of the sensor site. In orer to improve the etetion n lssifition ury while reuing the flse lrm rte, it is reommene tht the seismi n PIR sensor shoul be use together to provie omplementry informtion to eh other. Informtion fusion tehniques re neee to ombine the outputs of the two sensing molities, n this is topi of future reserh. Fiel eployment of sensors lrgely epens on the tsks n terrins. To enhne the perimeter seurity [7] in n open fiel, the sensors re usully eploye linerly or in irles. Sine the intruer my pproh the seure region from ny iretion, the worst se senrio is when the intruer pprohes the seure region extly hlf-wy between two jent sensors long stright pth perpeniulr to the sensor piket line. To ensure intruer etetion, the mximum sensor sping shoul be less thn the effetive rnge of the sensor. Therefore, sensor eployment oul be very expensive, beuse the etetion rnge of PIR sensors is less thn 6 m. Another ritil pplition is sensor eployment t hoke points; sine the trgets re fore to pss the hoke point ue to the terrin iffiulties, single noe of UGS system n be suffiient to over the entire region. V. CONCLUSION This pper presents symboli feture extrtion metho for trget etetion n lssifition, where the fetures re extrte s sttistil ptterns by symboli ynmi moeling of the wvelet oeffiients generte from time series of seismi n PIR sensors. By pproprite seletion of wvelet bsis n sle rnge, the wvelet-trnsforme signl is enoise reltive to the originl time-omin signl. In this wy, the symboli imges generte from wvelet oeffiients pture the signl hrteristis with lrger fielity thn those obtine iretly from the time omin signl. The symboli imges re then moele using probbilisti finite stte utomt (PFSA) tht, in turn, generte low-imensionl sttistil ptterns, lso lle feture vetors. A istint vntge of the propose feture extrtion metho is tht the low-imensionl feture vetors n be ompute in-situ n ommunite in rel time over limite-bnwith wireless sensor network with limite-memory noes. The propose metho hs been vlite on set of fiel t ollete from ifferent lotions on multiple ys. A omprtive evlution is performe on the seismi signls between SDF n kurtosis nlysis using leve-one-out rossvlition. Results show tht SDF hs superior performne over kurtosis nlysis, espeilly in the humn/niml lssifition. In ition, the pbilities for ientifying movement type n trget pylo re exmine for the seismi sensor. A three-wy ross-vlition hs been use to ssess the performne of PIR sensors for trget etetion n lssifition. Results show tht PIR sensors re very goo for trget etetion, n hs omprble performne with seismi sensors for trget lssifition n movement type ientifition. While there re mny reserh issues tht nee to resolve before exploring ommeril pplitions of the propose metho, the following topis re uner tive reserh: 1) Enhnement of trget etetion n lssifition performne by fusion of seismi n PIR sensor signls 2) Rel-time fiel implementtion of the propose metho on low-ost low-power miroproessors for ifferent types of eployment (e.g., UGS fening to seure region). REFERENCES [1] G. L. Grhm, Detetion n lssifition for unttene groun sensors, in Pro. Inform. De. Control, 1999, pp [2] D. Li, K. D. Wong, Y. H. Hu, n A. M. Syee, Detetion, lssifition, n trking of trgets, IEEE Signl Proess. Mg., vol. 19, no. 2, pp , Mr. 22. [3] A. G. Rihr, M. B. Bennett, n D. T. O Brien, Vehile n personnel etetion using seismi sensors, in Pro. Sens. C3I Inform. Trining Tehnol. Lw Enforement, Boston, MA, 1999, pp [4] J. Altmnn, Aousti n seismi signls of hevy militry vehiles for o-opertive verifition, J. Soun Vibrt., vol. 273, nos. 4 5, pp , 24.

11 1718 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 212 [5] Y. Tin n H. Qi, Trget etetion n lssifition using seismi signl proessing in unttene groun sensor systems, in Pro. Int. Conf. Aoust. Speeh Signl Proess., Orlno, FL, 22, p [6] G. P. Sui, D. Clpp, R. Gmpert, n G. Pro, Footstep etetion n trking, Pro. SPIE, vol. 4393, pp , Apr. 21. [7] J. Lombe, L. Pek, T. Anerson, n D. Fisk, Seismi etetion lgorithm n sensor eployment reommentions for perimeter seurity, Pro. SPIE, vol. 6231, p , Apr. 26. [8] M. Kenneth n P. D. MGffign, Spetrum nlysis tehniques for personnel etetion using seismi sensors, Pro. SPIE, vol. 59, pp , Sep. 23. [9] Y. Shoji, T. Tksuk, n H. Ysukw, Personl ientifition using footstep etetion, in Pro. Int. Symp. Intell. Signl Proess. Commun. Syst., 24, pp [1] L. Pek n J. Lombe, Seismi-bse personnel etetion, in Pro. IEEE 41st Ann. Int. Crnhn Conf. Seurity Tehnol., Ottw, ON, Cn, 27, pp [11] Z. Zhng, X. Go, J. Bisws, n K. K. Wu, Moving trgets etetion n loliztion in pssive infrre sensor networks, in Pro. 1th Int. Conf. Inform. Fusion, Jul. 27, pp [12] A. Ry, Symboli ynmi nlysis of omplex systems for nomly etetion, Signl Proess., vol. 84, no. 7, pp , 24. [13] V. Rjgopln n A. Ry, Symboli time series nlysis vi wveletbse prtitioning, Signl Proess., vol. 86, no. 11, pp , Nov. 26. [14] S. Gupt n A. Ry, Symboli ynmi filtering for t-riven pttern reognition, in Pttern Reognition: Theory Applition. Huppge, NY: Nov Siene Publishers, 27. [15] X. Jin, S. Gupt, K. Mukherjee, n A. Ry, Wvelet-bse feture extrtion using probbilisti finite stte utomt for pttern lssifition, Pttern Reognit., vol. 44, no. 7, pp , 211. [16] P. Abry, Onelettes et Turbulene, Multiŕesolutions, Algorithmes e Déomposition, Invrine Déhelles. Pris, Frne: Dierot Eiteur, [17] D. Lin n M. Mrus, An Introution to Symboli Dynmis n Coing. Cmbrige, U.K.: Cmbrige Univ. Press, [18] C. M. Bishop, Pttern Reognition n Mhine Lerning. NewYork: Springer-Verlg, 26. Xin Jin (S 9) reeive the B.S. egree in mehnil engineering from Shnghi Jio Tong University, Shnghi, Chin, in July 27, n the M.S. egree in eletril engineering n mehnil engineering from Pennsylvni Stte University, University Prk, where he is urrently pursuing the Ph.D. egree in mehnil engineering. His urrent reserh interests inlue mhine lerning, signl proessing, ontrol systems, energy mngement, n robotis. Mr. Jin is Stuent Member of the Amerin Soiety of Mehnil Engineers, the Interntionl Soiety of Automtion, n the Amerin Nuler Soiety. Soumly Srkr reeive the B.S. egree in mehnil engineering from Jvpur University, Kolkt, Ini, in 21. He is urrently pursuing the M.S. egree in mehnil engineering from Pennsylvni Stte University, University Prk. His urrent reserh interests inlue mhine lerning, sensor networks, n fult ignostis. Asok Ry (SM 83 F 2) reeive the Ph.D. egree in mehnil engineering from Northestern University, Boston, MA, n Grute egrees in the isiplines of eletril engineering, mthemtis, n omputer siene. He joine Pennsylvni Stte University (Penn Stte), University Prk, in July 1985, n is urrently Distinguishe Professor of Mehnil Engineering n Grute Fulty Member of eletril engineering. Prior to joining Penn Stte, he hel reserh n emi positions with the Msshusetts Institute of Tehnology, Cmbrige, n Crnegie-Mellon University, Pittsburgh, PA, s well s reserh n mngement positions with GTE Strtegi Systems Division, Wshington D.C., Chrles Strk Drper Lbortory, Cmbrige, n MITRE Corportion, Befor, MA. He hs uthore or o-uthore over 5 reserh publitions inluing 25 sholrly rtiles in referee journls n reserh monogrphs. Dr. Ry is Fellow of the Amerin Soiety of Mehnil Engineers n the Worl Innovtive Fountion. He h been Senior Reserh Fellow with the NASA Glenn Reserh Center uner Ntionl Aemy of Sienes Awr. Shlbh Gupt (S 3 M 6) reeive the M.S. egree in mehnil n eletril engineering from Pennsylvni Stte University (Penn Stte), University Prk, n the Ph.D. egree in mehnil engineering from Penn Stte in August 26. He joine the Deprtment of Eletril n Computer Engineering, University of Connetiut (UConn), Storrs, s n Assistnt Professor in August 211. Before joining UConn, he ws Post-Dotorl Reserh Sholr n s Reserh Assoite with the Mehnil Engineering Deprtment, Penn Stte. His reserh efforts hve been instrumentl in opening new fiels of t unerstning n pttern isovery, vi interfing multiisiplinry onepts erive from sttistil mehnis, symboli ynmis, n utomt theory. His urrent reserh interests inlue siene of utonomy, swrm robotis, intelligent systems, mhine lerning, network siene, n fult etetion n isoltion in omplex systems. Dr. Gupt is member of the Amerin Soiety of Mehnil Engineers. Thygrju Dmrl (M 84 SM 95) reeive the B.S. n M.S. egrees from the Inin Institute of Tehnology, Khrgpur, Ini, n the Ph.D. egree from Boston University, Boston, MA. He is urrently n Eletronis Engineer with the U.S. Army Reserh Lbortory, Aelphi, MD. He ws with the Inin Spe Reserh Orgniztion, Srihrikot, Ini, from 1973 to 1979, n the Inin Institute of Tehnology, Knpur, Ini, from 1979 to He ws n Assistnt Professor with the University of Kentuky, Lexington. He publishe more thn 1 tehnil ppers in vrious journl n onferenes n hs three U.S. ptents. His urrent reserh interests inlue signl proessing n sensor fusion for situtionl wreness.