Ieraoal Joural of Eergy, Iformao ad Commucaos, pp.9-20 hp://dx.do.org/10.14257/jec.2013.4.6.02 Objec Trackg Based o Ole Classfcao Boosed by Dscrmave Feaures Yehog Che 1 ad Pl Seog Park 2 1 Qlu Uversy of Techology, Cha 2 Uversy of Suwo, Korea cheyh@spu.edu.c, pspark@suwo.ac.kr Absrac We propose a robus objec rackg algorhm uder rackg by deeco framework. A sparse radom Gaussa-dsrbued measureme marx s used o buld a objec appearace model he compressed doma, ad deeco s doe by our sem-supervsed learg process. Our mehod ras a group of weaker classfers correspodg o every feaure, ad learly combes hem usg he weghs compued by s dscrmave power o buld a srog classfcao fuco, whch deecs a arge from backgroud. Kalma flerg s also adoped o smoohe he rackg resuls. Expermes show he proposed classfcao mproves performace rackg whou drf. Keywords: Objec rackg, Sparse radom measureme marx, Nave Bayes classfer, Dscrmave power, Feaure seleco, Kalma fler 1. Iroduco Objec rackg s a mpora ask wh he feld of compuer vso, ad s applcao area s broad. I s a challegg problem due o umerous facors lke appearace chages, vewpos of movg objecs, ad a cluered backgroud uder dffere llumao codos. Sgfca progress has bee made durg hese years. Trackg ad assocaed problems of feaure seleco, objec represeao, dyamc shape, ad moo esmao became very acve research areas ad ew soluos are beg proposed [14]. I wha s called "rackg by deeco", a objec s ofe represeed by a vecor a feaure space, ad a rackg ask s formulaed as a bary classfcao problem so ha each mage a vdeo sequece s classfed o eher a arge or a backgroud [12]. Ths kd of rackg algorhms coss of a few ma seps: feaure exraco, feaure seleco, classfer rag, ad deeco every frame. I realy, he represeao of a objec s a very hgh dmesoal space of complee feaures, causg he curse of dmesoaly. There are some mehods volved such a problem, ad sparse represeao obaed by L1 orm mmzao [3] s oe such example. Whe buldg a objec s appearace model, feaure exraco ad feaure seleco are wo ma ways drecly relaed o dmesoaly reduco. Recely he compressve sesg mehod has bee successfully roduced objec rackg [15], makg use of sparse represeao of a arge. We call her algorhm RCT for shor. Eve hough RCT s a ousadg poeerg work ha combes varous sae-of-he-ar echques a cocse maer, we foud ha ca be furher 2 Correspodg Auhor ISSN: 2093-9655 IJEIC Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos mproved by adopg a explc feaure seleco process ad a er-frame predco mechasm for robus objec rackg. I rece years, feaure seleco has bee aracve, ad a lo of work has bee doe [4, 7, 10, 11, 13]. The goal of feaure seleco s o pck a good subse of feaures from some large se of caddae feaures [6]. Feaure seleco plays a mpora role performace of classfcao. Our ma dea s o formulae feaure s dscrmave power from rag daa, ad ulzg o wegh exraced feaures. These feaure weghs ca be used o boos weaker classfers (correspodg o dvdual feaures) o a srog classfcao, whch replaces he classfer RCT. For er-frame predco, we adop he Kalma fler, whch also allows sgle-frame measureme of hdde saes boh for he poso of rackg objecs ad he feaure weghs. Ths paper s orgazed as follows: I Seco 2, we roduce some relaed works. I Seco 3, we roduce he fudameal prcples from whch our mehod s deduced. I Seco 4, expermeal resuls are preseed wh dscusso ad aalyss o how ad why our mehod works well, followed by cocluso ad fuure works. 2. Relaed Works 2.1. RCT The real-me compressve rackg (RCT) mehod by Zhag e al., [15] s a sae-ofhe-ar algorhm, erms of effcecy, accuracy ad robusess. The followg s a bref summary abou RCT [1, 2, 8, 15]. RCT s a rackg algorhm wh a appearace model based o he feaures exraced from a mul-scale mage feaure space wh a daa-depede bass. Is appearace model employs o-adapve radom projecos ha preserve he srucure of he mage feaure space of objecs. A rackg ask s formulaed as a bary classfcao va a ave Bayes classfer wh ole updae a compressed doma. The whole rackg process cosss of wo sages: learg ad deeco. A arge s maually labeled oly a he begg a gve vdeo, ad rag daa s assocaed auomacally wh he prevous rackg resul aferwards. I he learg sage, gve he capured arge by he prevous rackg resul, he algorhm crops rag samples (boh he arge ad he backgroud), aroud he arge s poso some predefed rego, whch are he used for learg/updag he classfer. I he ex frame, he deecor samples paches as pu o he classfer o look for he arge. Fally he algorhm defes he poso of he sample pach wh maxmum poseror as he poso of he objec. The he ere wo-sage rackg process s repeaed. 2.1.1. Feaure exraco ad dmesoaly reduco RCT m I vdeo rackg, each mage pach ca be represeed by a vecor x, ad le m X be a feaure space. A feaure dcoary s a bass of he feaure space X. The umber m of feaures he dcoary ca be huge, ad ay objec ca be represeed by a lear combao of feaure bass vecors, where m. RCT uses he feaure exraco approach based o compressve sesg by roducg a m sparse radom Gaussa measureme marx R o projec a hgh-dmesoal mage m vecor x oo a lower dmesoal oe v by v Rx (1) where he (, j) -h compoe of R s gve by 10 Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos rj 1 s 0 1 wh probably 1/2s wh probably 1-1/s, wh probably 1/2s RCT uses s=m/4, hece R s very sparse, requrg very low compuaoal complexy [1, 15]. Sce he ozero eres of R are eher 1 or -1, he compressve feaures compue he relave esy dfferece a way smlar o geeralzed Haar-lke feaures, whch s a lear combao of radomly geeraed recagles [2, 6, 15]. As a llusrao, see Fgure 1. Every group of recagles he same color correspods o a radom Haar-lke feaure. The fgure o he rgh s a geeral emplae for feaure exraco. Fgure 1. Feaures wh a sample frame. The move s from [16] 2.1.2. The ave Bayes classfer RCT For rackg by deeco, we eed a raed classfer ha ca dsgush wo classes of samples,.e., arge samples ad backgroud samples he feaure space. Le {( x, y ) : 1,2,... s, y {0,1}} be he rag daa, where each x s a rag sample m he se X, ad v s s low-dmesoal represeao. s s he umber of rag samples, ad y s a class label,.e., 0 for egave ad 1 for posve samples. RCT uses a ave Bayes classfer. Samples sasfy he followg codoal probably, p( v y 1) p( v 1) ( 1) 1 y p y. p( v y 0) p( v 0) ( 0) 1 y p y If p( v y 1) p( v y 0), v wll be labeled as posve, ad egave oherwse. A ave Bayes classfer uses a log rao o represe he classfer fuco. RCT selecs he sample wh he larges F(v) as he arge measureme, where p( v 1) ( 1) 1 y p y ( 1) ( ) log p v log y Fv p( v 0) ( 0) p( v 0) 1 y 1 y p y 2.2. Kalma fler The Kalma fler s a recursve esmaor ha uses a seres of measuremes observed over me, coag ose ad oher accuraces, ad produces esmaes of ukow varables ha ed o be more precse ha hose based o a sgle measureme aloe. More formally, he Kalma fler operaes recursvely o sreams of osy pu daa o produce a sascally opmal esmae of he uderlyg sysem sae. Objec rackg ca be hough of as a hdde Markov cha problem. Deeco by classfcao focuses o he curre frame oly, ad does o reflec he raslao of he objec bewee wo cosecuve frames. To deal wh lkg bewee wo (2) Copyrgh c 2013 SERSC 11
Ieraoal Joural of Eergy, Iformao ad Commucaos cosecuve frames, we adop Kalma flerg. Fgure 2 shows he role of he Kalma fler rackg [5, 7]. 3. Proposed Algorhm Fgure 2. Kalma flerg rackg by deeco I vsual objec rackg, he mos desrable propery of a feaure s s uqueess so ha he objec s easly dsgushable he feaure space [6, 14]. Furhermore, he feaures ha bes dscrmae bewee he arge ad he backgroud are desrable. A radoal way s choosg feaures maually, depedg o he applcao doma. I mache learg ad paer recogo commues, a wde rage of auomac feaure seleco algorhms have bee vesgaed,for example, AdaBoos s a popular algorhm wdely used mache learg commuy [11]. However, mos of hese mehods eed offle rag, huge rag daa, a log rag perod, ad all he formao kow advace. For objec rackg, ole seleco of dscrmave feaures s more suable due o he chage objec appearace frame by frame. Through he aalyss of he source code of RCT obaed from [16], we foud ha lacks a explc feaure seleco process accordg o he characerscs of specal objecs. Is feaure exraco process uses a off-le calculaed radom Gaussa measureme marx, keepg he same feaure emplaes hroughou he rackg ask. However we formulae feaure s dscrmave power from rag daa, ad use o wegh boh he releva ad rreleva feaures. These feaure weghs are he used o boos weaker classfers (.e., ave Bayes classfers) o a srog classfcao, whch ca replace he classfer RCT. We ra our classfcao model wo seps: frs, lear a group of weaker classfers f correspodg o every feaure (assumg v has a depede dsrbuo wh each oher) respecvely. The assg a wegh vecor o weaker classfers o boos o a srog classfcao. The job ca be doe by ay of feaure seleco mehods, bu here s some rade-off. We do o use oe of he off-he-shelf feaure seleco modules, bu formulae feaure s dscrmave power from rag daa o maa hgh effcecy ad a he same me o mprove rackg accuracy. Ths mehod may be hough of as a smple replaceme of he Fsher dscrma whch s used o evaluae feaure s dscrmave power classfer rag, bu our mehod s smpler ad easy o realze. 3.1. Weak classfers RCT reas all exraced feaures equally mpora whou cosderg he explc srucure of he arge. However o all feaures are of equal mporace. I our mehod, ulke (2), we roduce max so ha f( v ) 0. p( v 1) ( ) max{0,log y f v } p( v y 0) (3) 12 Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos f ca be used as a weaker classfer correspodg o every feaure. These wll be combed o form a srog classfcao, ad some rreleva feaures ca be removed by seg her weghs o 0. For a radom sample mages, we assume he uform pror, p( y 1) p( y 0). The radom projecos of hgh dmesoal vecors are kow o be almos always Gaussa dsrbued. Thus, he codoal dsrbuos p( v y 1) ad p( v y 0) f( v ) are assumed o be Gaussa dsrbued, wh four parameers (µ 1,σ 1, µ 0, σ 0 ), where µ 1 1 ad σ are he mea ad he sadard devao of he posve sample se X 1 he -h feaure, ad µ 0 ad σ 0 are he mea ad he sadard devao of he egave sample se X 0 he -h feaure,.e., p v y ~ N(µ 1,σ 1 ), ad p( v y 0) ~ N(µ 0,σ 0 ). ( 1) f( v ) s leared ad updaed ole, accordg o he rag sample ses X 1 ad X 0. 3.2. Srog classfcao boosed by dscrmave power of feaures Afer learg a group of weaker classfers f (v ) correspodg o each feaure, we eed o ge a wegh vecor W o buld a srog classfcao as a lear combao of hem. We wa o use a smple ad drec way o fd W=(W 1,,W ) ad ole boos weaker classfers o a srog classfcao. Feaures wh more dscrmave power should have hgher weghs cosrucg a srog classfcao. The defo of dscrmave power ca vary depedg o applcaos. Sce posve samples wll look more smlar o he shape of he arge ha egave samples, f some feaure s dscrmave eough, he dsrbuo of feaure values of posve samples wll be more coceraed ha ha of egave samples. Fgures 3.A) ad 3.B) show examples of some good dscrmave feaures. The blue curve s for posve samples ad he red oe s for egave samples, ad hey show sgfca dfferece. O he oher had, he wo curves of o-dscrmave feaures do o show such dfferece, as show Fgures 3.C) ad 3.D). 1 We may coclude ha good releva feaures wll have smaller σ bu larger σ 0, ad he bgger he dfferece bewee he wo, he greaer he dscrmave power of he correspodg feaure. Hece we defe a measure of dscrmave power by w, ad he form a wegh W afer ormalzao : W w 1 w 0 1 where w 0 1. (4) Fgure 3 shows he probably dsrbuos of 4 dffere feaures - he blue les for posve samples ad he red oes for egave samples. I geeral, feaures wh bgger dfferece he wo sadard devaos wll have more dscrmave power. For a desrable feaure, he blue curve should show a more coceraed paer ha ha of he red curve (Fgures 3.A) ad 3.B). The oppose examples are 3.C) ad 3.D). Hece, we roduced he measure w (4). Our expermeal resuls show ha he assumpo s plausble. Throughou hs paper, we use he superscrp o deoe frame umbers. Le W be he wegh of each weak classfer a frame, ad defe a srog classfcao by 1 1 1 F( v) W f ( v ) (5) Copyrgh c 2013 SERSC 13
.. Wegh vecor Ieraoal Joural of Eergy, Iformao ad Commucaos Ths classfcao model s ole refreshed, ad s beefcal o defy he ma characerscs uder osy codo. I (5), weaker classfers f are boosed by he wegh vecor W o a srog classfcao F. A) B) C) D) Fgure 3. The dsrbuo of four feaure values. The frs wo have more dscrmave power ha he oher wo Our algorhm may look smlar o geeral boosg algorhms, bu here s a sgfca dfferece. Mos boosg algorhms have eravely learg weak classfers wh respec o daa dsrbuo ad add hem o form a srog classfer. Whe weak classfers are added, hey are ypcally weghed some way ha s usually relaed o weak learer s accuracy. The erao s geerally expeced o coverge fally, or sops a a prese erao upper boud. As a cosequece, he umber of weak classfers s o cera. However, our mehod, weak classfers are weghed some way ha s relaed o feaure s dscrmave power raher ha her accuracy. I addo, here s o erao. The advaage of our mehod s ha s smple ad quck eough so ha ca be used for ole rag. Fgure 4 s he schemac dagram of our srog classfcao. The red doed box shows he dfferece from ha of RCT. μ 1 Feaures evaluag ad weghg Posve samples Feaure Exraco learg σ 1 μ 0 σ 0 Classfer parameers Classfer parameers Tme -1 Negave samples Feaure Exraco Classfer parameers Tme Ipu daa Weak classfer... Weak classfer... Weak classfer Deecor samples Feaure Exraco Srog classfcao maxmum a poseror deeco resul Fgure 4. A schemac dagram of our srog classfcao 14 Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos 3.3. Trackg resuls A he -h frame, le he prevous sample respose be z -1 =(x,y,w,h), where x -1 =(x,y) s s coordae, ad w ad h are s wdh ad hegh, respecvely. Le W -1 be he wegh he prevous frame. Le d be he umber of deecor samples, ad V { V1,..., Vj : j 1,..., d} be he represeao of curre deecor samples whch are ceered a he po wh a cera searchg radus from he pror poso. The ex sample respose wll be ake as he oe wh maxmal value of F: J arg max F(V j ), j=1,2,,d. (6) j I deeco sage, he algorhm samples paches wh a predefed search rego. The he classfer looks for he arge by lookg for he sample pach wh maxmum poseror. 3.4. Smoohg by Kalma fler Objec rackg ca be hough of as a hdde Markov cha problem. Deeco by classfcao geeral oly focuses o he curre frame, ad does o reflec he raslao bewee wo cosecuve frames. So far, we dd o cosder he moveme of a arge objec. However o deal wh lkg bewee frames, we ow adop a Kalma fler. I our rackg applcao, hdde saes are he objec s poso ad feaure s wegh, ad he Kalma fler wll calculae he esmaed saes. Our Kalma fler bul o a smple moo model s as follows: 1 W I I 0 0 W W 0 I 0 0 W, 1 1 x 0 0 Im m I mm x 1 x 0 0 0 I mm x W W 0 0 0 K I W x 0 0 Im m 0 K x x ad are oses whch obey Gaussa dsrbuo. W s he wegh vecor ad x s he objec s acual poso. We alze he wegh of each feaure b by seg W 0 =1/ ad W 0 =0 for all. x 0 s maually alzed he frs frame ad se Δx 0 =0. x K ad W K are he measuremes obaed from he deeco ad leared classfer. Sce he poso updae ad he wegh updae ca be doe depedely, we use wo dffere Kalma flers, say, KalmaP for poso ad KalmaF for wegh, respecvely. 3.5. Our algorhm Summg up, he flow of our full algorhm s show Fgure 5. The deal of he algorhm s show Algorhms 1, 2, ad 3 below. Copyrgh c 2013 SERSC 15
Ieraoal Joural of Eergy, Iformao ad Commucaos Fgure 5. Overall flow of our algorhm Algorhm 1. Ole learg ad updae Ipu: -h vdeo frame ad he prevous racked locao x Oupu: weaker classfer parameer vecor p, wegh vecor W 1. Sample wo ses of mage paches,.e., a posve sample se X 1 ={x x x α, =1,,r} ad a egave sample se X 0 ={x j β x j - x γ, =1,,r } where ad j are dces of paches, α, β, γ (α< β< γ) are appropraely chose parameers, ad r s he umber of mage paches. 2. Exrac feaures V wh low dmesoaly usg (1). 3. Lear he parameers p =(µ 1,σ 1, µ 0,σ 0 ) for each weaker classfer f (v ). 4. Lear he wegh vecor W k by (4). 5. To correc W k, use he Kalma fler o oba W = KalmaF(W k,δw ) where Δ W =W k W -1. Algorhm 2. Deeco Ipu: (+1)-h frame, weaker classfer parameers p, ad boosed wegh vecor W Oupu : x +1 1. Sample a se of mage paches D δ ={ x x -x <δ} where x s he rackg locao a he -h frame, ad exrac feaures V wh low dmesoaly. 2. Use he boosed classfcao F o each sample vecor V ad fd he rackg locao x +1 wh he maxmal classfcao respose by (6). 3. Use he Kalma fler o ge correcve poso by x +1 =KalmaP(x +1, x) where x = x +1 - x. Algorhm 3. Trackg by deeco Ipu: al arge poso x 0 ad Δx 0 =0. Oupu: rackg resuls 1. Ialze he Kalma fler wh W 0 =1/ ad W 0 =0 for all. 2. Repea from he frs frame ul he las frame, 16 Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos 2.1) Execue Algorhm 1 (Ole Learg ad updae). 2.2) Execue Algorhm 2 (Deeco). I our full algorhm, we apply he Kalma fler boh o he poso of he objec ad o he wegh. However, o esmae he effec of he Kalma fler, we devse wo varas of our full algorhm, depedg o where we apply he Kalma fler. They are summerzed as follows: SgKK : Our full algorhm. Use boh KalmaP ad KalmaF. (=SgKPF) SgKP : A vara usg KalmaP oly. SgKF : A vara usg KalmaF oly 4. Expermes ad Resuls We compare he resul of RCT wh hose of our full algorhm (sgkk) ad s varas (SgKP ad SgKF) usg he publc daase Davd door from [16]. We used he same parameers (.e., α=4, β=5, γ=20 Algorhms 1 ad 2). Oher parameers lke δ or he wdh ad hegh of he pach sampled are decded by maually markg he objec a he begg of he ru.) as ha he auhors used [15] excep for he umber of feaures, whch we creased from 50 o 100. We maually marked he exac poso of he objec ad drew a boudg box o compare wh he resuls by he above mehods. Fgure 6 shows he comparso. The whe boxes are he boudg boxes aroud he acual posos, whle he blue oes are by RCT, he gree oes by sgkp, ad he red oes by sgkk. As we ca see Fgure 6, red oes are he closes o whe boxes whle he blue oes by RCT are he farhes. Red - sgkk Whe - acual poso Blue - RCT Gree - sgkp Fgure 6. Comparso of rackg resuls by dffere algorhms Fgure 7 shows he esmaed dsaces bewee he acual poso ad ha obaed by each mehod. The x-coordae correspods o he frame dex, ad he y-coordae correspods o he dsace ( pxels) bewee he rackg resul by each mehod ad he acual poso. Ul frame 50, sgkk (=SgKPF) looks somewha usable compared o ha by RCT (=CT). However he log ru, sgkk works obvously beer ha RCT. We also draw wo resuls by sgkf ad sgkp o see he effec of KalmaF ad KalmaP. Ieresgly, he graph of sgkf s smlar o ha of RCT, mprovg oly very lle. O he oher had, eve hough sgkp somemes gves worse resuls ha RCT, eveually gves a beer resul ad KalmaP seems o be more mpora ha KalmaF. Ayhow hey seem o show some syergy effec sgkk. Copyrgh c 2013 SERSC 17
Ieraoal Joural of Eergy, Iformao ad Commucaos Fgure 7. Comparso of he rackg resuls by dffere algorhms AdaBoos s a feaure seleco algorhm wdely used mache learg commuy, ad seems feasble o mprove he performace of RCT by roducg AdaBoos. We obaed he AdaBoos code from [9], ad creaed a ole compressve AdaBoos racker (we ame OCA for shor) by replacg RCT s aïve Bayes classfer. Fgure 8 shows he comparso of he hree algorhms, OCA ( gree), RCT ( blue), ad sgkk ( red). The four small fgures he lower par show magfed vews for frames 1~100, 100~200, 200~300 ad 300~426. Ul frame 100, he performace of AdaBoos racker s he bes, whle ha of sgkk s he poores. Bu afer frame 100, sgkk shows he bes performace ul he ed. Probably oe of he reasos why AdaBoos shows poor performace afer frame 100 s, because he perso he vdeo ured hs head slghly, ad AdaBoos dd o have eough daa for rag as he shape of he objec chages. Fgure 8. Comparso of he hree algorhms usg he daa se [16] - OCA (gree), RCT (=CT,blue), ad sgkk (red) 5. Cocluso ad Fuure Work I hs paper, we proposed a rackg algorhm based o classfcao volvg ole feaure seleco. Our algorhm combes compressed represeao ad feaure seleco. Such a combao has bee rarely used oher leraures. We roduced a feaure valuao ad weghg mehod o combe weaker classfers o form a srog classfcao. Ths feaure seleco s depede of applcaos, ad ca be combed wh ay represeao of a feaure space o ra classfers. We also roduced he Kalma fler o deal wh he raslao of objec bewee wo cosecuve frames. Expermeal resuls show ha our assumpo s feasble ad he formulao based o esmao of dscrmave feaures s prove o be effce. 18 Copyrgh c 2013 SERSC
Ieraoal Joural of Eergy, Iformao ad Commucaos To realze successful loger erm rackg, seems ha we eed o perform more elaborae aalyss o rak each feaure s dscrmave power for a beer dscrmave feaure bass. Oe of he drecos mgh be employg he Fsher dscrma aalyss o rak each feaure s dscrmave power ad combg wh he AdaBoos algorhm. Our frs ry wh AdaBoos was o successful, showg a beer performace oly a he begg. We wll focus o how o make AdaBoos keep such a good performace log erm rackg. Noeheless, here always exss a radeoff bewee accuracy ad effcecy, whch we eed o balace o our way. Ackowledgemes The auhors would lke o hak Dr. K. Zhag for makg her precous vdeos publcly avalable wh o resrco. They also hak Professor A. L a Qlu Uversy of Techology Cha for makg helpful dscusso o her work. Ths work was suppored by he GRRC program of Gyeogg provce [GRRC SUWON2013-B1, Cooperave CCTV Image Based Coex-Aware Process Techology]. Refereces [1] D. Achlopas, Daabase-fredly radom projecos: Johso-Ldesrauss wh bary cos, J. Compu. Sys. Sc, vol. 66, o. 4, (2003), pp.671 687, do: 10.1016/S0022-0000(03)00025-4. [2] B. Babeko, M. -H. Yag ad S. Beloge, Robus objec rackg wh ole mulple sace learg, IEEE Tra. J. Paer Aalyss ad Mache Iellgece, vol. 33, (2011) Augus, pp. 1619-1632. [3] C. Bao, Y. Wu, H. Lg ad H. J, Real Tme Robus L1 Tracker Usg Acceleraed Proxmal Grade Approach, IEEE CVPR, (2012), pp. 1830 1837. [4] R. Colls, Y. Lu ad M. Leordeau, O-Le Seleco of Dscrmave Trackg Feaures, IEEE Trasaco J. Paer Aalyss ad Mache Iellgece, vol. 27, o. 10, (2005) Ocober, pp.1631 1643. [5] E. Cuevas, D. Zaldvar ad R. Rojas, Kalma Fler for Vso Trackg Techcal Repor B, (2005). [6] P. Dollar, Z. Tu, H. Tao ad S. Beloge, Feaure mg for mage classfcao, IEEE Coferece o CVPR 07, (2007) Jue, pp. 1-8. [7] Z. Ha, Q. Ye ad J. Jao, Ole Feaure Evaluao for Objec Trackg Usg Kalma Fler, Paer Recogo, ICPR, (2008), pp. 1-4. [8] Z. Kalal, J. Maas ad K. Mkolajczyk, P-N Learg: Boosrappg Bary Classfers by Srucural Cosras, CVPR, (2010). [9] D. -J. Kroo, AdaBoos algorhm, hp://www.mahworks.com/malabceral/fleexchage/27813-classc- AdaBoos-classfer. [10] D. W. Lag, Q. M. Huag, W. Gao ad H. X. Yao, Ole Seleco of Dscrmave Feaures Usg Bayes Error Rae for Vsual Trackg, 7h Pacfc-Rm Coferece o Mulmeda, (2006), pp. 547 555. [11] K. Okuma, A. Talegha, N. D. Freas, J. J. Lle ad D. G. Lowe, Boosed Parcle Fler: Mularge Deeco ad Trackg, ECCV04, Par I, (2004) May, pp. 28-39, do: 10.1007/978-3-540-24670-1_3. [12] S. Salder, H. Graber ad L. va Gool, Beyod Sem-Supervsed Trackg: Trackg Should Be as Smple as Deeco, bu o Smpler ha Recogo, IEEE ICCV Workshops, (2009) Ocober, pp. 1409 1416. [13] J. Wag, X. Che ad W. Gao, Ole Selecg Dscrmave Trackg Feaures Usg Parcle Fler, Proceedgs of IEEE CVPR, vol. 2, (2005), pp. 1037 1042. [14] A. Ylmaz, O. Javed ad M. Shah, Objec rackg: A survey, J. ACM Compug Surveys (CSUR), vol. 38, Issue 4, Arcle o.13, (2006), do: 10.1145/1177352.1177355. [15] K. Zhag, L. Zhag ad M. Yag, Real-Tme Compressve Trackg, ECCV, (2012). K. Zhag, L. Zhag ad M. Yag, Real-me Compressve Trackg, hp://www4.comp.polyu.edu.hk/~cslzhag/ct/ct.hm. Copyrgh c 2013 SERSC 19
Ieraoal Joural of Eergy, Iformao ad Commucaos Auhors Ye H. Che s a lecurer a Iformao School of Qlu Uversy of Techology, Cha. Her research eres cludes compuer vso, mache learg, mage daa mg, ad algorhm desg. She receved her M.S. degree Lgh Idusry Machery from Shaax Uversy of Scece ad Techology, Cha. Pl S. Park s a professor of Compuer Scece ad he drecor of Iformao & Compuer Ceer a Uversy of Suwo, Korea. Hs research eres cludes compuer vso, especally rackg objecs for survellace, hgh performace compug, kowledge-based formao sysems, ad Lux clusers. He receved hs Ph.D. Ierdscplary Appled Mahemacs (wh emphass o Compuer Scece) from Uversy of Marylad a College Park, M.S. Appled Mahemacs from Old Domo Uversy, U.S.A., ad B.S. Physcal Oceaography from Seoul Naoal Uversy, Korea. 20 Copyrgh c 2013 SERSC