Overlapped Vehicle Tracking via Enhancement of Particle Filter with Adaptive Resampling Algorithm

Transcription

1 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... Overlapped Vehile Traking via Enhanemen of Parile Filer wih Adapive Resampling Algorihm Wei Leong Khong, Wei Yeang Kow, Loria Angeline, Ismail Saad, Kenneh Tze Kin Teo Modelling, Simulaion and Compuing Laboraory, Shool of Engineering and Informaion Tehnology Universii Malaysia Sabah, Koa Kinabalu, Malaysia. mslab@ums.edu.my, kkeo@ieee.org Absra - Traffi surveillane and on-road seuriy have elevaed he demand of mahine vision aided raffi onrol sysem. Through he modern video amera ehnology, vehile raking has beome a vial approah o assis he on-road raffi sysems. In he pas, many raking mehods have been developed based on he deail and informaion exraed from he apured vehile. However, onvenional raking sysem need o be improved sine he bakground noises and sudden appear objes will inrease he diffiulies of oninuously raking he arge vehile. Hene, a parile filer algorihm wih adapive resampling approah has been proposed o overome he vehile olusion problems. In addiion, he proposed resampling approah an also be used o solve he ommon parile degeneray problem. Experimenal resuls show ha he enhaned parile filer equipped wih adapive resampling algorihm is signifianly improving he auray of he raking proess wihou ompromising he proessing ime. Keywords - parile filer, adapive re-sampling, vehile raking, raffi surveillane I. ITRODUCTIO II. REVIEWS OF OBJECT TRACKIG Vehile raking is an essenial approah ha has drawn he aenion among he researhers due o is numerous fields of appliaions suh as road raffi onrol sysem, raffi surveillane and seuriy sysem [1]. However, olusion and overlapping beween vehiles is a hallenging ask in surveillane sysem via image proessing. Due o he diffiulies and omplexiy aused by he olusion problems, he researhers are inied o sudy he effeive and effiien vehile raking mehod. In his researh, video sensor is hosen as he raking infrasruure raher han ohers sensors beause of he rapid developmen in video amera ehnology. Furhermore, a wide range of he informaion o desribe he arge vehile suh as he olour, moion, edge, shape and speed of he vehile an be exraed from he video sensor via image proessing ehniques. Alernaively, vehile flow onsiss of dynami hanges whih ould lead o non-linear and non-gaussian ondiions. Hene, parile filer has been hosen as he vehile raking algorihm in his researh due o is abiliy o overome he non-linear and non-gaussian siuaions. everheless, parile degeneray was he main faor ha will influene he auray of vehile raking resuls. Therefore, an effiien and effeive resampling approah will be needed o solve he ommon parile degeneray problem. Thus, an enhanemen of he parile filer wih he adapive resampling sep is implemened o oninuously raking he arge vehile under various overlapping inidens wihou ompromising he proessing ime. In he pas, many mehods have been developed for obje raking purpose. Among he well known ehniques of image proessing raking mehods are Markov hain Mone Carlo [2], Kalman filer [3, 4], opial flow and parile filer [5, 6, 7]. Vehile flow onsiss of dynami hanges may lead o non-linear and non-gaussian siuaion. In his ase, he developmen of exended version of Kalman filer an be used o overome he non-linear siuaion. However, when he nonlineariy is inauraely approximaed by he exended Kalman filer ehnique, he esimaed resuls will be diverging and hene lead o an inaurae raking resul. Parile filer is proven as a promising and powerful ehnique o overome he nonlinear and non-gaussian siuaion. I has been hosen as he overlapped vehile raking ehnique due o is abiliy o ope wih he non-linear and non-gaussian sae [8, 9]. In researh [10], he non-rigid objes have been raked by using he olour feaure. I is suggesed ha he algorihm seleed for obje raking purpose should be able o deal wih he parial olusion and sale invarian inidens. everheless, olour is a powerful feaure ha an be implemened in hese siuaions. The exraed olour hisogram of he arge vehile will be omparing wih he olour hisogram of he sample vehile by using Bhaaharyya disane. As disussed in [10], he olour based algorihm an effiienly handle he non-rigid and fas moving objes under differen ondiions. Aording o referene [11], he ypial parile filer will fae a phenomenon named as parile degeneray during he raking proess. Parile filer degeneray ours due o he low weigh or weak parile is seleed 44 ISS: x online, prin

2 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... afer several ieraions and i bloks he furher improvemen of he algorihm. In general, here are wo ways used in solving he parile degeneray problem. The soluion is eiher inreasing he number of parile size implemened in he algorihm or resampling he pariles. However, inrease he parile size is insuffiien due o he huge amoun of he sample size ould lead o higher ompuaional omplexiy and proessing ime. On he oher hand, resampling an inrease he auray of he raking resuls by eliminaing he low weigh pariles and regenerae wih srong weigh pariles wihou ompromising he proessing ime [12]. As menioned in [13], olusion and sudden appear objes iniden are among he hallenging asks ha will be faed in vehile raking. Therefore, he proposed parile filer algorihm mus be robus wheher in he parially or fully olusion inidens. When he arge obje is being overlapped, he informaion of he arge obje will be vanished or influened by he obsale. The raking performane based on he resuls is aurae bu he algorihm fails when here is a large degree of disjoin during he raking. III. METHODOLOGY This seion disusses he mehodology of he parile filer and explains he likelihood of he samples being ompued. A. Parile Filer Framework In his seion, a brief review of parile filer will be presened. Parile filers also known as sequenial Mone Carlo whih is an ieraive proess ha esimaes poserior disribuion from a finie se of weighed pariles [14]. I is developed based on esimaing he urren sae of he arge vehile from he previous parile se. Basially, he onvenional parile filer algorihm onsiss of hree imporan seps whih are prediion sage, measuremen sage and resampling sage. In he prediion sage, he parile filer will generae a new pariles se wih eah parile represens he esimaed poserior posiion. The inremen of he number of pariles an improve he auray of he esimaion. Meanwhile, he ompuaional ime will be longer wih he large amoun of he parile size. In he measuremen sage, eah parile weigh is ompued based on likelihood probabiliy. For visual raking purpose, he observaion sae of he arge vehile an be olour, edge or shape ha exraed from he informaion of he arge vehile. In his sudy, olour feaure has been seleed as he parameer for he vehile raking. The hird sage refers o he resampling proess. Resampling is an imporan sep o redue he parile degeneray problems. Parile degeneray affes he auray of he raking resuls. During parile degeneray, he low weigh parile is oninually seleed and used by he parile filer algorihm whih leads o inauray resuls. Hene, resampling algorihm is needed o eliminae he low weigh parile and regenerae a new se of pariles unil he parile ses wih large weigh is obained. I is imporan o avoid he parile degeneray problem o improve he auray of he raking resuls. The dynami hanges in vehile raking usually onsis of nonlinear and non-gaussian elemens. The poserior probabiliy densiy funion as desribed in (1) an be obained hrough he prediion sage. However, he observaion probabiliy densiy funion as in (2) is used o express he likelihood of he olour feaure. p X Z ) (1) ( p Z X ) (2) ( The quaniies of he raked obje is denoes by sae veor X while all he observaions sae a ime is denoes by veor Z. In he prediion sage, he prior probabiliy densiy funion an be obained hrough (3) and he poserior probabiliy densiy funion is defined using he Bayes rule in (4). p (3) ( X Z1 : 1) p( X X 1) p( X 1 Z1: 1) dx 1 p( Z X ) P( X Z1: 1) p ( X Z1: ) (4) p( Z Z ) 1: 1 In he parile filer algorihm, he poserior probabiliy densiy funion developed from he prior densiy is represened by a se of weighed parile samples. Meanwhile, poserior densiy funion an be obained hrough (5) beause he weighed pariles are in disree naure and w is he normalized weigh as shown in (6). i i p( X Z1 : ) w ( X X ( i)) (5) i1 i i i i i p( z x ) p( x x 1) w w 1 (6) q( x x, z ) 1: 1 B. Color Disribuion Model In his researh, olour feaure has been hosen for he vehile raking purpose beause of is abiliy o deal wih he parial olusion and sale invarian problems. Moreover, he proessing ime o obain he olour informaion of he arge vehile is muh faser han oher 1: 45 ISS: x online, prin

3 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... parameers. Hene, olour feaure has been hosen and implemened in mos of he visual raker. The olour hisogram of he arge vehile is normally alulaed in he RGB olour spae o obain a disree bins hisogram. Afer obaining he olour hisogram of he arge vehile, i will be ompared wih he olour hisogram of he referene vehile in order o ompue he similariy of he wo hisograms whih is alled likelihood. Bhaaharyya disane is a ommon ehnique used o measure he likelihood beween wo olour hisograms. Afer he obaining he Bhaaharyya disane, he weigh of he parile an be alulaed based on he likelihood. C. Bhaaharyya Disane Generally, he measuremen beween wo olour disribuion hisogram is alulaed by he ommon ehnique alled Bhaaharyya oeffiien [15]. In his ase, he Bhaaharyya oeffiien is used o alulae he oeffiien beween olour hisogram of referene vehile, } and olour hisogram of he arge vehile, p { p u u1... q { q u } u1... as shown in (7). [ p, q] p q du (7) Sine he olour hisogram is a disree densiy model, he Bhaaharyya oeffiien an be obained hrough (8). u1 u u [ p, q] p q (8) The value of Bhaaharyya oeffiien represens he similariies of wo olour disribuion. The larger value of he oeffiien means more similariy in he olour disribuion. However, he limi of he oeffiien is se from 0 o 1. If boh he hisogram is idenial, hen he oeffiien will be indiaed as 1. Afer obaining he Bhaaharyya oeffiien, he Bhaaharyya disane an be ompued using (9). b dis u u 1 [ p, q] (9) Based on he Bhaaharyya disane, he weigh of he pariles an be alulaed using (10) where σ is he adjusable sandard deviaion whih an be hosen experimenally. 2 b dis 2 2 e 2 1 (10) 2 The weigh of he pariles is se o heavy when he olour of he referene vehile and he olour of he arge vehile are oo similar. This esimaed posiion represened by he heavy parile will beome he possible loaion of he arge vehile and i will be updaed in he measuremen sage of he parile filer algorihm. IV. RESAMPLIG ALGORITHM As disussed in he previous seion, resampling proess is needed o eliminae he parile degeneray problem. This seion will desribe he parile degeneray in more deail. Then he onvenional resampling and proposed resampling algorihms are disussed owards he parile degeneray siuaion. A. Parile Degeneray Parile filer is a good approah in vehile raking due o he abiliy o deal wih non-linear and non-gaussian siuaions. However, afer several ieraions, he parile filers also faing he problem aused by parile degeneray. Generally, implemening he parile filer algorihm wih large size of parile samples or resampling he parile samples is apable of avoiding he parile degeneray problem. Alhough boh approahes an be used o improve he auray of he vehile raking, resampling is more suiable o be implemened ino he parile filer algorihm. Resampling is hosen sine he ompuaional ime is muh lesser ompared o applying huge number of pariles. In order o measure he appearane of he parile degeneray problem, he effeive sample sizes need o be alulaed using (11). Sine, he rue weigh of pariles in (12) anno be deermined, an esimae of he effeive sample size will be ompued using (13). where w eff * i eff s (11) *i 1Var( w ) i p( x z1: ) (12) i i q( x x, z ) s i1 1 1 ( w ) i 2 (13) i The w in (13) is alulaed using (6) and i is he normalized weigh of eah parile. Afer obaining he esimaed effeive sample size, (13) is used o indiae wheher parile degeneray problem ours or no. 46 ISS: x online, prin

4 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... If eff s, he parile degeneray problem has ourred and resampling is needed. The parile filer will keep on reursively resampling unil he requiremen is reahed. B. Convenional Resampling Resampling is ommonly used o overome he parile degeneray problem. In he pas, various ype of resampling algorihm has been developed. The mos ommon resampling algorihms being used o redue he parile degeneray are mulinomial resampling, residue resampling, sysemai resampling and sraified resampling. In Table I, eff is he esimaed effeive sample size whih is used o deermine wheher he parile degeneray ours or no. When he parile degeneray ours during he vehile raking proess, he esimaed effeive sample size will be less han he iniiaed sample size. As a resul, resampling will be aivaed and a new se of parile samples is regeneraed. The newly generaed parile samples will be reweighed via he olour likelihood ehnique. If he new se of he parile samples is no able o reah he hreshold of he effeive sample size, he resampling proess is hen oninually being evaluaed unil he requiremen is fulfilled. When he overlapping vehile is raked by using he radiional resampling algorihm, he raking resuls will no be promising. This is beause when he arge vehile is being oluded by he obsale, he informaion ha an be exraed by parile filer will be limied. Therefore, more ieraion of resampling algorihm is needed o obain an aurae raking resul. However, afer a few ieraion of resampling proess, he informaion of he vehile will be influened by he obsale and hene he arge vehile will be los rak. Besides ha, more ompuaion ime is onsumed if he parile filer is reursively repeaing he resampling sage due o he dissimilariy olour hisogram of he referene vehile and he olour hisogram of he arge vehile. C. Proposed Resampling In his sudy, an enhaned parile filer algorihm wih he adapive funion of resampling is proposed. Wih he proposed resampling algorihm, he resampling ompuaional ime will be opimized in vehile raking under various overlapping ondiions. The enhaned resampling approah only resamples he low weigh pariles where he weigh is below he hreshold value. Meanwhile, he pariles wih he aeped weigh will be sored as he referene posiion for vehile raking purpose. This is o shoren he resampling proess and hene redue he ompuaional ime wih a promising vehile raking resuls. In shor, i is suiable o be used TABLE I. COVETIOAL RESAMPLIG ALGORITHM PF wih Convenional Resampling Algorihm 1: Iniialize referene olour hisogram and sample size 2: FOR FRAME = 1, 2,, 3: PREDICTIO: 4: FOR i = 1, 2,, s 5: Draw predied pariles from prior dynamis 6: Compue he olour hisogram based on esimaed posiion 7: ED FOR 8: MEASUREMET: 9: Calulae he Bhaaharyya disane, b dis 10: Compue he weigh of he parile based on Bhaaharyya disane, i i i 1 11: ormalize he weigh, w w ( 1 w ) 12: Calulae eff 13: WHILE eff s 14: RESAMPLIG: 15: Resample he disree disribuion 16: Generae new se of pariles 17: Compue he weigh of he parile based on Bhaaharyya disane, 18: Chek he effeive sample size 19: ED WHILE 20: OUTPUT: 21: Obain he posiion of he arge vehile 22: ED FOR FRAME when he vehile is wihou olusion or overlapping. However, when he arge vehile is being oluded, he informaion exraed by he parile filer algorihm migh be inorre and hene i migh wrongly rak he arge vehile due o lak of resampling proess. Thus, during he olusion, only he larges weigh of he parile will be reserved for resampling purpose. Based on he larges weigh of he parile, i is fas o rak he arge vehile even only a small porion of he vehile s olor is visible. Moreover, he proposed resampling algorihm an also assis in gaining bak he informaion of he arge vehile faser han oher resampling algorihm. The enhaned resampling algorihm for vehile raking is shown in Table II. Meanwhile, he flowhar of he parile filer wih he adapive resampling algorihm is illusraed in Fig. 1. V. TARGET LOCALIZATIO In his seion, esimaion of he loaion for he arge vehile will be disussed. Afer he enire pariles are weighed by he likelihood, he parile filer algorihm will 47 ISS: x online, prin

5 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... eff s Oupu of esimaion End O Sar Iniialize sample size Predi poserior posiion of pariles Calulae Bhaaharyya Disane Compue weigh of parile ormalized he parile weigh Deermine effeive sample size YES eff T hres YES Choose he highes weigh parile Regenerae pariles based on highes weigh parile posiion Figure 1. Flowhar of parile filer wih adapive resampling algorihm. deermine he needs of he resampling sep. If resampling sep is exeued, he algorihm will hek for he olusion iniden via he hreshold. When he esimaed effeive sample size is lower han he hreshold, he arge vehile is parially oluded or fully oluded. Hene, he highes weigh of he parile will be remained as he referene posiion for he nex prediion sep while eliminaing ohers pariles. This aion an speed up he algorihm o gain bak he informaion of he arge vehile when here are a small porion olour of he arge vehile is visible. On he oher hand, when he esimaed effeive sample size is higher value han he hreshold, he arge vehile is no oluded. Thus, in he resampling sep, he parile filer only eliminae he low weigh parile and a he same ime he heavy weigh parile will be remained. Afer ha, he number of eliminaed parile will be resampled and his sep is repeaing unil he esimaed effeive sample size fulfils he requiremens. The enhaned resampling proess shorens he ompuaional ime ompared o he onvenional resampling algorihm beause i only resamples he pariles whih poenially ause he degeneray problem. Afer resampling, oupu sage exiss as he esimaed posiion of he arge vehile. The posiion of he arge vehile an be easily alulaed hrough he mean of all he oordinaion generaed by he pariles. This is beause afer he resampling sage, all he pariles will be foused and onenraed a he enre of he arge vehile. O Eliminae he low weigh parile Regenerae he remaining of he low weigh parile TABLE II. ADAPTIVE RESAMPLIG ALGORITHM PF wih Adapive Resampling Algorihm 1: Iniialize referene olour hisogram and sample size 2: FOR FRAME = 1, 2,, 3: PREDICTIO: 4: FOR i = 1, 2,, s 5: Draw predied pariles from prior dynamis 6: Compue he olour hisogram based on esimaed posiion 7: ED FOR 8: MEASUREMET: 9: Calulae he Bhaaharyya disane, b dis 10: Compue he weigh of he parile based on Bhaaharyya disane, i i i 1 11: ormalize he weigh, w w ( 1 w ) 12: Calulae eff 13: WHILE eff s 14: RESAMPLIG: 15: IF eff Thres 16: Choose he highes weigh of he parile 17: Resample he disree disribuion based on highes weigh parile 18: ELSE 19: Eliminae all low weigh pariles and remain he good parile 20: Regenerae he remaining of he pariles 21: ED ELSE 22: Compue he weigh of he parile based on Bhaaharyya disane, 23: Chek he effeive sample size 24: ED WHILE 25: OUTPUT: 26: Obain he posiion of he arge vehile 27: ED FOR FRAME VI. RESULTS AD DISCUSSIOS In his seion, he resuls of vehile raking using he onvenional parile filer resampling algorihm as shown in Fig. 2 will be ompared o he resuls of vehile raking using an enhaned parile filer resampling algorihm as shown in Fig. 3. In boh resampling algorihm, he parile size is iniialized as 200 samples. Colour hisogram is seleed as he parameer o alulae he weigh of he pariles. A olour hisogram wih bins RGB olour spae will be ompued. In his sudy, he olour feaure has been hosen due o is abiliy o idenify he 48 ISS: x online, prin

6 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... arge vehile ideniy during parial olusion. In addiion, when here are a small porion of he olour of he arge vehile is visible, he enhaned parile filer algorihm is able o loae he arge vehile. Sine olour hisogram in disree form, he ime required o proess will be shor. Based on he resuls shown in Fig. 2 and Fig. 3, he rossing ion represens he loaion esimaed by he parile filer algorihm. Meanwhile he solid box refers he bounding box of he arge vehile. The arge vehile posiion is deermined by alulaing he mean value of he oordinaes esimaed by eah parile. Referring o Fig. 2 and Fig. 3, i an be observed ha he raking resuls an be divided ino four ases whih are before oluded, parially oluded, fully oluded and afer oluded. Comparing o he resuls shown in Fig. 2 and Fig. 3, i an be noied ha he resuls obained by using he parile filer wih adapive resampling algorihm is muh more promising. In Case I, he arge vehile is no oluded by anoher sai vehile as shown in Frame 5 of Fig. 2 and Fig. 3. Based on he resuls, he onvenional and enhaned parile filer resampling algorihms are able o rak he arge vehile. This is beause before he olusion, he informaion of he arge vehile an be easily obained wihou influenes by anoher vehile. Thus, he parile filer resamples as usual by eliminaing he weak pariles and replaing hose unwaned pariles wih a new se of pariles. In Case II, he arge vehile is parially oluded by he sai vehile as shown in Frame 13 of Fig. 2 and Fig. 3. From he resul shown, i is noied ha olour is an imporan feaure ha an be used o deal wih parially olusion inidens. In his ase, he resampling algorihm is exeued as he previous ase beause he informaion of he arge vehile sill obainable and no influened by he sai vehile. Thus, he arge vehile is able o be raked wih boh resampling algorihms. In Case III, he arge vehile is almos fully oluded by he sai vehile as shown in Frame 16 of Fig. 2 and Fig. 3. Mos of he informaion of he arge vehile is eiher los or influened by he informaion obained for he sai vehile. Thus, he onvenional resampling algorihm is failed o loae he arge vehile beause he pariles has been rapped a he sai vehile by geing he wrong informaion. Meanwhile, he improved resampling algorihm is able o rak he arge vehile alhough only a small porion of he arge vehile olour is reognized. In Case IV, he arge vehile ours afer he olusion as shown in Frame 25 of Fig. 2 and Fig. 3. Based on he resuls obained, he onvenional resampling algorihm is unable o rak he arge vehile. The onvenional resampling algorihm failed in raking he vehile afer olusion is beause he informaion of he arge vehile has been los and replaed by he informaion of he sai vehile. Therefore, he arge vehile is onsidered los rak by using he onvenional parile filer algorihm as Frame 5 Frame 13 Frame 16 Frame 20 Frame 25 Figure 2. Resuls of vehile raking by using onvenional resampling parile filer. Frame 5 Frame 13 Frame 16 Frame 20 Frame 25 Figure 3. Resuls of vehile raking by using adapive resampling parile filer. 49 ISS: x online, prin

7 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... shown in Frame 25 of Fig. 2. However, he enhaned parile filer wih adapive resampling is able o oninuously rak he arge vehile as shown in Frame 25 of Fig. 3. Alhough he informaion of he arge vehile will be influened by he sai vehile afer he overlapping, he improved resampling algorihm is apable of remaining he loaion of he highes weigh parile. A new se of samples will be generaed o gain bak he informaion of he arge vehile. Hene, he arge vehile is able o be oninuously raked by he proposed algorihm effeively and effiienly. Fig. 4 and Fig. 5 show he RMSE and he resampling proess as well as he sample size versus he frame index. Based on Fig. 4 and Fig. 5, he frame index of 1 o 10 indiaes he arge vehile is free from olusion. The resuls shows ha boh of he resampling algorihms are able o rak he arge vehile due o he low value of RMSE. However, in erms of ompuaional ime, he adapive resampling is faser han he onvenional resampling as shown in Fig. 5 due o he lesser resampling seps and number of resampling pariles. During he frame index of 11 o 14, he arge vehile is parially oluded by he sai vehile. From he resul shown in Fig. 4, he RMSE for boh resampling algorihms are almos he same. Afer frame index of 14, he number of resampled pariles has been inreased due o he influees from he sai vehile as shown in Fig. 5. Frame index of 16 shows ha he arge vehile is fully oluded by he sai vehile. From Fig. 4, he RMSE for he onvenional algorihm is muh more higher han he proposed algorihm. This is beause he informaion of he arge vehile is influened by he sai vehile and auses he onvenional algorihm o an inaurae resul. Meanwhile, he raking resul for he enhaned resampling algorihm is sill promissing wih he low value of RMSE. In addiion, Fig. 5 shows ha he onvenional algorihm will keep on resampling o loae he arge vehile and beome an infiniy loop. Thus, a maximum of 20 resampling seps has been se in order o erminae he infiniy loop. Frame index of 20 o 25 indiaes he arge vehile is ourred afer he olusion. From he resuls shown in Fig. 4, he RMSE of he onvenional algorihm is inreasing due o he algorihm is sill suk a he sai vehile. The inreasing of he RMSE means ha he parile filer has diverged from raking he arge vehile. However, RMSE for he enhaned algorihm is sill remain a low level whih means he visual raker is sill oninuosly rak he arge vehile. On he oher hand, he number of resampling seps required for ha adapive algorihm is sill mainain a a low level. This means he adapive algorihm an gain bak he informaion of he arge vehile during and afer he overlapping. However, he onvenional algorihm is failed in raking he arge vehile even hough wih he maximum of he resampling seps as shown in Fig. 5. The ompuaional ime aken by he Figure 4. Graph of RMSE vs frame index. Figure 5. Graph of resampling oun and resampling sample size vs frame index. adapive algorihm is shorer han he onvenional algorihm due o he number of resampling seps and resampled parile sizes is lesser. VII. COCLUSIO As disussed in he previous seions, he auray of he parile filer algorihm will be diminished by he parile degeneray. Conepually, resampling an be used o avoid he parile degeneray problem. Thus, an enhanemen of he parile filer wih adapive resampling algorihm has been proposed for he purpose of raking he overlapped vehile. The implemenaion of he 50 ISS: x online, prin

8 WEI LEOG KHOG e al: OVERLAPPED VEHICLE TRACKIG VIA EHACEMET OF... adapive algorihm in he parile filer approah is apable o ease he raking diffiulies of various olusion inidens. The performane and robusness of he proposed algorihm is esed and assessed under various raking ondiions. I an be onluded ha he adapive resampling approah has improved he auray of he raking resuls wihou ompromising he ompuaional ime. ACKOWLEDGEMET The auhors would like o aknowledge he finanial assisane of he Minisry of Higher Eduaion of Malaysia (MoHE) under Fundamenal Researh Gran Shemes (FRGS) o. FRG0220-TK-1/2010 and he Universiy Posgraduae Researh Sholarship Sheme (PGD) by Minisry of Siene, Tehnology and Innovaion of Malaysia (MOSTI). REFERECES [1] F. Gusafsson, F. Gunnarsson,. Bergman, U. Forsell, J. Jansson, R. Karlsson and P.J. ordlund. Parile Filers for Posiioning, avigaion, and Traking. IEEE Transaion on Signal Proessing, vol. 50, no. 2., pp , [2] W.Y. Kow, W.L. Khong, F. Wong, I. Saad and K.T.K. Teo. Adapive Traking of Overlapping Vehiles via Markov Chain Mone Carlo wih CUSUM Pah Plo Algorihm. In Proeedings of Compuaional Inelligene, Communiaion Sysems and eworks, 2011, pp [3] P.L.M. Bouefroy, A. Bouzerdoum, S.L. Phung and A. Beghdadi. Vehile Traking using Projeive Parile Filer. In Proeedings of 6 h IEEE Inernaional Conferene on Advaned Video and Signal Based Surveillane, 2009, pp [4] J. Liu and P. Vadakkepa. Ineraing MCMC Parile Filer for Traking Maneuvering Targe. Digial Signal Proessing, vol. 20, pp , [5] J. Li, X. Lu, L. Ding and H. Lu. Moving Targe Traking via Parile Filer Based on Colour and Conour Feaures. In Proeedings of 2nd Inernaional Conferene on Informaion Engineering and Compuer Siene, 2010, pp.1-4. [6] J. Czyz, B. Risi and B. Maq. A Color-based Parile Filer for Join Deeion and Traking of Muliple Objes. In Proeedings of IEEE Inernaional Conferene on Aousis, Speeh, and Signal Proessing. 2005, pp [7] C.J. Yang, R. Duraiswami and L. Davis. Fas Muliple Obje Traking via a Hierarhial Parile Filer. In Proeedings of 10 h IEEE Inernaional Conferene on Compuer Vision, 2005, pp [8] M.S. Arulampalam, S. Maskell,. Gordon and T. Clapp. A Tuorial on Parile Filer for Online onlinear/on-gaussian Bayesian Traking. IEEE Transaion on Signal Proessing, vol. 50, no.2, pp , [9] H.P. Liu, F.C. Sun, L.P. Yu and K.Z. He. Vehile Traking using Sohasi Fusion-based Parile Filer. In Proeedings of IEEE/RSJ Inernaional Conferene on Inelligen Robos and Sysems, 2007, pp [10] K. ummiaro, E. Koller-meier and L.V. Gool. Colour Feaures for Traking on-rigid Objes. Speial Issue on Video Surveillane Chinese Jornal of Auomaion, vol. 29, pp , [11] T. Wada, F. Huang and S. Lin. Visual Traking Using Parile Filers wih Gaussian Progress Regression. Springer-Verlah Berlin Heidelberg PSIVT 2009, LCS 5414, 2009, pp [12] X.Y. Fu and Y.M. Jia. An Improvemen on Resampling Algorihm of Parile Filer. IEEE Transaion on Signal Proessing, vol. 58, no.10, pp , [13] A.D. Bagdanov, A.D. Bimbo, F. Dini and W. umziai. Adapive Unerainy Esimaion for Parile Filer-based Trakers. In Proeedings of 14 h Inernaional Conferene on Image Analysis and Proessing, 2007, pp [14] T. Zhang, S.M. Fei, X.D. Li and H. Lu. An Improved Parile Filer for Traking Color Obje. In Proeedings of Inernaional Conferene on Inelligen Compuaion Tehnology and Auomaion, 2008, pp [15] M.S. Khalid, M.U. Ilyas, M.S. Sarfaraz and M.A. Ajaz. Bhaaharyya Coeffiien in Correlaion of Gray-Sale Objes. Journal of Mulimedia, vol.1 no. 1, pp , ISS: x online, prin