IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. Humn Body Trcking wih Auxiliry Mesuremens Mun Wi Lee, Isc Cohen Insiue for Roboics nd Inelligen Sysems Inegred Medi Sysems Cener Universiy of Souhern Cliforni Los Angeles, CA 90089-073 {munlee icohen}@usc.edu Absrc This pper presens wo echniques for improving humn body rcking wihin he pricle filering scheme. Boh echniques explore he use of uxiliry mesuremens. The firs echnique uses opicl flow cues o improve he smpling disribuion. The second echnique involves he deecion of individul body prs, nmely he hnd, hed nd orso; nd using hese deecion resuls o provide ddiionl inference on subses of se prmeers. This mehod enbles he uomic iniilizion of se vecor nd llows recovering from rcking filures. These wo mehods improve he overll ccurcy, efficiency nd robusness of humn body rcking s illusred by he experimenl resuls. Inroducion Humn body rcking is n cive reserch re wih mny pplicions in inercive virul environmen, humn compuer inercion, moion cpure for humn nimion nd video surveillnce (see [3] for survey. There re mny mehods for humn body rcking, including blob rcking [], grdien mehods [4][] nd voxel reconsrucion []. Recenly, he pricle filer (.k.. Condension [7], sequenil Mone Crlo hs been widely used [5][6] [0][8][9]. Bsed on smpling pproximion nd likelihood compuion, pricle filer is ble o hndle pose mbiguiies due o moion singulriies nd occlusion [5]. There re however some problems in he use of pricle filer. Firsly, in humn body rcking, he se vecor is of lrge dimension (ypiclly 30 resuling in inefficien smpling nd high compuion complexiy. Secondly, n ccure pose iniilizion is required prior o he rcking. Ofen, iniilizion is done mnully nd herefore reduces he prcicliy of he mehod in mny pplicions.. Reled Work Recen works using pricle filer for humn body rcking hve focused on improving efficiency using vrious mehods such s covrince scled smpling [9], simuled nneling pproch [6], priioned smpling [0] nd hybrid Mone Crlo pproch [3]. Oher works include combining pricle filer wih deerminisic grdien descen serch [], nd using dynmic moion model [8][8]. However, he lrge compuion complexiy remins n imporn issue for body rcking using pricle filer. Deecions of hed, fce, hnd nd orso hve been used in surveillnce nd hnd gesure pplicions. While here re problems wih view-poin vriion nd occlusion, he deecion resuls re generlly relible in mjoriy of viewing condiions. These body pr deecion mehods hve been used o provide corse esimes of he humn pose [][] nd for humn deecion in sic imges [4]. However hese mehods hve no been used for deiled humn body rcking becuse of he difficuly in deecing immedie joins such s elbows, shoulders, knees nd hips. Moion feures cn lso be used for performing beer rcking nd reducing he numericl complexiy of he pricle filer. Opicl flow hs been used for deecion nd rcking problems [3][8]. I provides dense field h is useful for inferring movemen of he subjec nd esiming he body joins displcemens. In our work, we consider body pr deecion nd moion feures s uxiliry mesuremens h re useful for inferring some specs of he humn pose nd moion. Our conribuions re new echniques for inegring hese mesuremens o he sequenil Mone Crlo scheme for more robus nd efficien rcking. In his work, we re using hree clibred sionry cmers.. Our Approch Our pproch is bsed on he use of uxiliry mesuremen o improve he se esimion of pricle
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. filer. These uxiliry mesuremens include moion cue nd body pr deecion. Here, he moion cue is used o improve he smpling disribuion by providing beer esimions of he curren posiions of he upper nd lower rms. The second ype of uxiliry mesuremen is he deecion of hnds, hed nd orso. The resuls of hese deecions re used o mke inference on subse of se prmeers corresponding o he observed humn body pose. This ddiionl inference is used o improve he se esimion wihin he pricle filering. The use of uxiliry mesuremens reduces he degree of rndomness during se vecor propgion nd lso reduces he relince on Mone Crlo simulion, hereby improving efficiency. The use of uxiliry mesuremen lso generes relible hypoheses on join posiions h help in iniilizing he pose nd recovering from rcking filure. The proposed pproch is fully uomic no requiring he iniilizion of he riculed body model nd moreover, he lgorihm boosrps iself llowing he recovery of body joins fer loss of rcking. The pper is orgnized s follows: Secion briefly describes he humn model nd he likelihood mesure considered. Secion 3 describes he use of moion cues. Secion 4 describes he use of body prs deecion. Experimenl resuls re presened nd discussed in Secion 5. Pricle Filer for Ariculed Humn Body Model Fiing The use of pricle filer for humn body rcking hs been well described in previous works [5][6][0][8]. Among hese works, here re some vriions in he implemenion, such s he smpling sregies nd he likelihood funcions. In his secion, we will briefly describe he riculed humn body model we use nd he likelihood funcion considered. The echniques proposed in his pper re generlly independen of he likelihood compuion nd re lso pplicble in conjuncion wih oher smpling sregies such s nneling pricle filer nd priioned smpling. Humn Body Model The riculed humn body model consiss of 0 joins nd 4 segmens, represening he hed, orso nd limbs. Ech segmen is represened by pered 3D cone wih n ellipicl cross-secion (Figure. The model hs 3 degrees of freedom h include he globl scle, rnslion, roion nd locl join roions (Tble. These degrees of freedom re represened in he se vecor x. Figure. The riculed humn body model considered in his work consiss of 0 joins nd 4 segmens. Prmeers Nme Globl Scle Globl Trnslion Globl Roion Join Angles Number 3 3 5 Tble. Model Prmeers Likelihood Compuion Fiing nd rcking he riculed model o he deeced humn in he video srems requires he definiion of likelihood funcion llowing he mpping of he degree of freedom of he model ono imge properies. The likelihood compuion is bsed on wo ypes of feures, he silhouee boundry nd silhouee regions. Given prediced pose (described by se vecor of pricle, we synhesize silhouee ono ech cmer view. This synhesized silhouee llows he mching of he esimed pose o he perceived silhouee derived by foreground deecion mehod. The likelihood mesure is compued by mching he boundries nd regions wih he erced foreground boundry. Furher deils cn be found in []. 3 Moion Cue for Improved Smpling Disribuion The 3D dynmic se vecor consiss of he globl scle, posiion nd orienion of he humn nd he join ngles. The se vecor chrcerizes he pose of he person ime. The se rnsiion model is defined by: x = x +η, ( where η is vecor of sndrd norml rndom vribles represening he process noise. This ssume zero order moion model nd h he humn is generlly sionry bu wih some rndom perurbions. This is n
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. engineering rick s i is difficul o model he humn dynmic chrcerisics ccurely in n unconrolled environmen. A he predicion sge of he pricle filer, (i he smpling of he ih pricle, x, employs proposl disribuion h is Gussin cenered on he previous ( se ~ i, nd defined by: x ( ~ ( i (i N x,σ, ( where Σ is he covrince mrix of he process noise. When he person is moving, he use of his proposl disribuion is no suible since he process noise will hve lrge vrince. However, wihou ddiionl informion, he previous se is he opiml unbised esimion. An lernive pproch consiss in using firs or higher order moion model. For exmple, firs order models re commonly used for rcking wlking person. However, he pplicion we re focusing is where he person is gesuring freely nd we do no expec body join ngles o chnge in consisen wy over resonble period of ime. In ddiion, incresing he order of moion model inroduces ddiionl prmeers nd will ffec rcking sbiliy nd increse he compuion ime. Anoher lernive is o model humn dynmics nd pply dimensionliy reducion echnique o reduce complexiy. This is useful for rcking resricive se of humn moion (e.g. wlking. For pplicions where here is higher vribiliy in humn movemen, he benefi of his pproch is uncler, nd lrge se of rining d would be required for lerning he dynmic model. In he bsic pricle filer mehod, he pricles propge using he rnsiion prior nd do no ke ino ccoun curren observion d, unil ler sge when he likelihood mesure is compued. This leds o inefficien smpling. Vrious pproches hve been suggesed o ddress his issue. The unscened pricle filer (UPF [], uses he curren observion o derive n improved proposl disribuion. Similr sregies re used in [8][7][5] for objec rcking. However, for humn body rcking, i is difficul o pply UPF direcly becuse i requires good grouping nd lbeling of edge feures which usully involves serching nd mching of erced feures wih synhesized feures obined by projecing he esimed body model ono he ll cmer views. We hve doped differen sregy bsed on using moion feures. This is chieved by considering opicl flow field o esime he moion of individul body prs nd genering beer proposl disribuion for he dynmic se prmeers. Since opicl flow is dense field, moion inference cn be redily obined wihou he need of serching nd lbeling. We define n uxiliry prmeer vecor (h we will refer o s moion prmeers s he chnge in se vecor v x x (3 = nd he se rnsiion model becomes x = + v (4 The chnge in se prmeer will induce opicl flow in he imges h cn be mesured. We define z s mesuremen of moion prmeers, so h z = v η, (5 where η is he mesuremen noise, nd he se rnsiion becomes x = x + z +η. (6 If he mesuremen, z, (which will be described ler is resonbly ccure, hen he mesuremen noise η will hve smll vrince compred o he process noise η. We pproxime he mesuremen noise η by vecor of sndrd norml rndom vribles: N 0,Σ. η ( We hen modify he proposl disribuion s, (i x ( ~ ( i ( i N x +, Σ z. (7 (i We noe h he mesuremen z is condiioned on he ( previous se ~ i becuse he posiions of he body prs re required before we cn erc he relevn moion prmeers from he opicl field. In he following, we describe his in more deils. For he presen discussion, we focus on he ercion of moion prmeers for he lower rms in order o erc he chnge of join ngle he elbow. For ech ( pricle, given is previous se ~ x i, we cn deermine he posiion of he lower rm on he imge. We inegre he opicl flow long he upper nd lower edges of he lower rm nd esime he D imge moion of he hnd wih respec o he elbow join. We only consider moion in he direcion perpendiculr o he lower rm in he imge. From ech imge, we cn only obin n esime of he moion projeced ono he plne orhogonl o he opicl ry. This inroduces wo consrins: combining he esimes from hree differen imges resuls in n over-consrin problem. The 3D moion of he lower rm is hen esimed using Les Men Squre mehod. The chnge in he join ngle of he elbow is hen esimed using he geomery of he riculed model.
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. ( (b Figure Mesuremen of moion prmeers lower nd upper rm. ( Opicl flow field, (b inegrion of flow long prediced edges of body prs nd esimion of moion long he norml direcion. In our implemenion, we use he pproch described bove o esime he 3D moion of ech of he lower nd upper rms nd upde he join ngles he elbows nd shoulders. This pproch is priculrly useful for inferring he rm movemen, becuse of he following resons:. In indoor environmen, while inercing wih sysem he movemens of he rms re usully fser nd more vrying compred o he res of he body. Therefore he join ngles he elbow nd shoulder hve lrger noise vrinces in he rnsiion model, reducing he efficiency nd he robusness during he rcking.. The movemens of he rms re usully observble from les wo of he cmers, so h resonbly good esimes of he rm movemen cn be erced from he opicl flow field. 4 Prs Deecion for Pril Upding of Se Vecor The second echnique involves he deecion of body prs o improve rcking by upding subses of he se prmeers, wihin he pricle filering scheme. In our implemenion, his scheme consiss of hree sges, where ech sge involves he deecion of specific body pr, nmely he hnd, hed nd orso respecively. The deecion resuls re used o upde subses of se vecor using inverse kinemics. The se prmeers h re upded ech sge re given in Tble. The ddiionl inference provided by he deecion of body pr improves he se esimion. The se esimion process for pricle is shown schemiclly in Figure 3. Deecion y Number of Consrins Upded prmeers Hnd 3 Elbow join (dof Shoulder join (dof. Hed 3 Neck join (dofs Neck posiion (dof Torso 4 Body posiion (dofs Body orienion (dofs Tble. Subse of he se vecor upded wih he body pr deecion. The proposed echnique is reled o Ro- Blckwellised pricle filering (RBPF [4], which improves efficiency by mrginlizing some of he se prmeers nlyiclly. Mos previous work in humn body rcking, uses edge nd region feures s mesuremen, from which i is difficul o compue ny priculr subse of he se vecor direcly. Therefore, RBPF is no redily pplicble o his problem wihou using differen ypes of mesuremen. Our proposed echnique, doping he RBPF pproch, inroduces he use of body prs deecion o mesure he posiions of specific body prs nd herefore is ble o mrginlize subses of he se vecor. In he following secions, we firs describe he mehods for deecing he body prs nd how hey re used o upde subses of se vecor. In he ler pr of his secion, we describe how his inference is inegred ino he overll se esimion frmework. Inpu Imge Figure 3. Body prs deecion nd se upding. Hnd Deecion Body Prs Deecion Se Prior Esimion (Pricle Filer y w x, x, x For hnd deecion, we loce cndide posiions s he peks of convex curvure long he oulines of he silhouee (Figure 4. We mch hese curvure peks in differen imges using epipolr consrins, nd reconsruc heir 3D posiions. x y x Se Upding Wih Inference x Likelihood Compuion
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. The mesuremen (giving 3 consrins is used o upde he posiions of hnd nd elbow, while keeping he res of he body posiion unchnged. Given he posiions of he hnd nd shoulder, he elbow lies long circle. We choose he poin on his circle h is closes o he previous elbow posiion. The join ngles he elbow nd shoulder re hen upded ccordingly. If cndide hnd posiion violes he physicl consrins of he body (e.g. oo fr wy from he shoulder, his cndide will be regrded s n oulier nd ignored. Hed Deecion The hed deecion is performed using reference chin code represenion of hed-shoulder conour (Figure 4b s emple for he hed. We mch his emple long he conour boundry of he erced silhouee o deec he locion of he hed. To chieve scle invrince, he conours re rescled wih respec o he esimed humn heigh. The chin code feures re normlized before comprison o chieve roion invrince. Posiions long he silhouee boundry h mch well re cndide for hed posiions. To obin he hed posiion in 3D spce, he hed mus be deeced in les wo inpu imges. Epipolr consrin is used o remove he flse mesuremens. A mesuremen of he hed posiion in he 3D spce provides hree consrins for upding he se prmeers. We choose o upde he hree degrees of freedom h re reled o he hed posiion: he orienion of he hed ( dofs nd he posiion of he neck long he body min xis ( dof. Chnge in he neck posiion necessrily generes chnge of he posiions of orso nd oher body prs such s he shoulders nd hips. Our sregy is o minimize chnges in he posiions of he end-effecors such s he hnds nd legs. We shif he orso only long is xis nd compue he join ngles he inermedie joins (shoulders, elbows, hips nd knees while keeping he hnds nd legs fixed. Torso Deecion A simple mehod is used o erc he min xis of he orso. We firs erc he medil xis of he D silhouees. The medil xis poins in differen views re mched using epipolr consrin, nd he 3D posiions re compued. A line is hen fied o hese 3D poins using PCA nd RANSAC mehods, bsed on les squre crierion. This erced line, illusred in Figure 4d, provides mesuremen of he orso orienion, nd consrin h he orso mus ly long he line. This gives four consrins, which re used o upde he se prmeers. The posiion ( dofs nd orienion ( dofs of he humn model re upded so h he orso is ligned o he erced medil xis (Noe h he posiion of orso long he xis, nd he roion round he xis remin unchnged. During his modificion, we pply he sme sregy s for hed deecion o keep he posiions of hnds nd legs consn nd upde he ngles inermedie joins. Deecion Errors The body pr deecions hve problems wih occlusions, s well s oulier errors, which will led o negive or flse deecion. To overcome his, no ll pricles re upded wih he deecion resuls. Insed, we use he likelihood mesure of he deecion (bsed on mching error during he mesuremen sges o deermine he confidence of he deecion resul. Mone Crlo mehod is hen used o decide wheher o use he deecion resul for ech pricle propgion. When he deecion resul is no used, he se vecor will be propged ccording o he sochsic se rnsiion model used in he sndrd pricle filer lgorihm. ( (b (c (d Figure 4. Body prs deecion. ( Hnd deecion, (b Temple of he hed. (c deeced hed, (d orso deecion, he medil xis of he silhouee in wo views is erced, mched nd reconsruced in 3D. The orso xis is found by fiing line o he medil xis poins in 3D. Inegrion wih Pricle Filer In he following, we describe he frmework on how he inference provided by body prs deecion is
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. inegred ino he pricle-filering scheme. Denoing x s he hidden se nd y s he observion. A ny ime, he se esimion is given by: p (, y y, (8 where p ( x x is he rnsiion probbiliy disribuion. In his secion, we consider only he se propgion for one pricle. Suppose h he se vecor x is decomposed ino wo prs ( x, x nd denoing y s n uxiliry mesuremen (from body prs deecion h cn be used o esime x nlyiclly. We cn rewrie he esimion expression s,,,, y, y y,,,, y y, y,,,,, y (9 For exmple, during he hnd deecion sge, he decomposiion of he se vecor is such h x consiss of he elbow join ngle nd wo join ngles he shoulder, which cn be inferred using inverse kinemics s described erlier. The oher subse x consiss of he remining 9 se prmeers. Since y is mesuremen on locl body pr (e.g. deecion of lef hnd, he res of he se prmeers in x (which rele o oher prs of he body, e.g. righ rms, legs ec re relively independen of y. We cn pproxime he esimion expression s:,,, y, y. y, y,,,, (0 In he simples cse, we cn compue x deerminisiclly by funcion x = f ( y,,,, so h: y,,, = δ ( f ( y,,, ( where δ ( is he Dirc del funcion. Then he esimion becomes: p (,,, y, y y x, x x x, x δ ( x f ( y, x, x, x ( This llows Mone Crlo smpling o be pplied only o he reduced se vecor x, herefore improving he efficiency. Muliple Hnd Deecion Cndides For he hnd deecion, we my hve deeced number of cndides for hnd posiion. In his cse, he inference ( y,,, p is non-zero for smll finie se of discree vlues. For ech pricle, we smple x from his finie se by Mone Crlo mehod, using he confidence mesures of ech cndide posiion (bsed on mching error. In his scenrio, he smpling disribuion hs collpsed from coninuous x spce o few discree vlues. While he inference is no olly deerminisic, he degree of rndomness hs been reduced considerbly. In his work, we re ssuming h he mesuremen noises re negligible. This is resonble ssumpion for hed nd hnd deecions h generlly hve good loclizion properies. In our curren work, we re ddressing he issue of mesuremen noise nd re reformuling he inference in probbilisic frmework, king ino ccoun he mesuremen uncerinies. 5 Experimenl Resuls 5. Experimen Seup nd Trcking Iniilizion Three clibred cmers re se up o cpure video sequences in room. A bckground of he empy scene is firs lerned for deecion purposes. As he person eners he views, he silhouees in he hree views re erced using bckground subrcion mehod. Once he person hs fully enered he scene, he pricle filer is iniilized using he body prs deecion mehods nd ddiionl heurisic rules described in []. The iniilizion is fully uomic nd does no require he person o snd in priculr posure. The firs iniilizion my no be ccure due o self-occlusions nd pose singulriies. Bu he use of uxiliry mesuremens is ble o genere ddiionl good se hypoheses o help in recovering more ccure pose subsequen frmes. Deils of his iniilizion sep cn be found in []. 5. Improvemen from Moion Cue In Secion 3, we describe he use of moion cue o modify he proposl disribuion for pricle smpling, so ( ( h he disribuion is cenered on ~ i i x + z, where ~ ( x i is he previous se fer re-smpling, nd is he esimed chnge in prmeers from moion cue. Wihou he use of moion cue, he disribuion will be cenered on ~ ( i. In his experimen, for ech ime frme, we evlue he likelihood mesures he wo ceners of disribuions using he curren imge, i.e. we compue nd compre ~ ( i ( i p ( y + z nd ~ ( i ( y x, where ~ i is he previous z (i
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. se prmeers of he bes mched pricle ime -. The resul is shown in Figure 5, which shows h he likelihood mesure is generlly higher wih he improved proposl disribuion genered using he moion cue. Log Likelihood -80-0 -60 Wihou Moion Cue Wih Moion Cue compuion in body prs deecion. Body prs deecion is performed only once for ech frme. The min compuionlly inensive processing is he likelihood compuion which is performed for every pricle ech frme. Boh mehods use he sme likelihood compuion. Figure 7 shows resul from noher video of person who is urning her body nd wving her hnds. The roion of he orso is no esily deeced from D silhouee edges, bu using he cues provided by he limbs, he rcker is ble o esime he pose correcly s shown by he rendered humn model. -00 Time Figure 5. Likelihood mesures wih nd wihou he use of moion cue. Noe h he likelihood is expressed in he logrihmic scle. 5.3 Recovering Trcking Filure nd Robus Trcking As only hree cmers re used, pose mbiguiies will occur while observing person gesuring due o selfocclusion, moion singulriies nd bckground cluer. The rcking mehod should be robus gins hese problems. While he sndrd pricle filer uses muliple pricles o smple he poserior disribuion of he se spce, i suffers from he problem of high dimensionliy, which cuses smple depleion in mos of he se spce. As resul, when n mbiguiy occurs, i is esy o loose rck nd recovering from los rcking is difficul. Our mehod uses body pr deecion o infer some of he se prmeers nd is ble o genere good se hypoheses, even in smple-depleed se spce region. This provides n venue for minining good rcking nd recovering from los rcking. We compre he performnce of our mehod wih he sndrd pricle filer. Boh lgorihms were esed wih sequence h conins brief period of bou 0 seconds when he hnds of he person were hidden behind his bck nd were occluded from ll cmers (Figure 6. Afer he hnds repper, he sndrd pricle filer is unble o recover from he los rck. In he proposed mehod, he riculed body model is ble o djus o he correc hnd posiions fer he person s hnds repper. In his experimen, 00 pricles re used for he new proposed mehod, while 400 pricles were required by he sndrd pricle filer lgorihm. In he proposed mehod, he reducion in compuion due o he smller number of pricles hs more hn offse he ddiionl ( (b (c (d Figure 6: Trck recovery fer self-occlusion. Firs row shows he resul from sndrd pricle filer wihou inference. Second row shows he resul from improved pricle filer wih inference. For ech row: ( before occlusion, (b boh hnds re occluded behind he person body, (c one hnd reppers, (d boh hnds re visible. Figure 7. Trcking urning person: The person is urning while wiving her hnds. 6 Conclusion We hve proposed wo improvemens o he pricle filer echnique for humn body rcking by using wo ypes of uxiliry mesuremens. Firsly, moion cue is used o improve he smpling disribuion by providing
IEEE Inernionl Workshop on Anlysis nd Modeling of Fces nd Gesures, 003. beer esimions of he curren posiions of he upper nd lower rms. Secondly, uxiliry mesuremen from he deecion of hnds, hed nd orso, re used o infer subses of se prmeers nd improve he esimion of humn pose wihin he pricle-filering scheme. This hs he dvnges of performing uomic rck iniilizion, recovering los rck, nd reducing he compuionl complexiy of he pricle filering mehod. Acknowledgemen The reserch hs been funded in pr by he Inegred Medi Sysems Cener, Nionl Science Foundion Engineering Reserch Cener, nd Cooperive Agreemen No. EEC-9595. 7 References [] Mun Wi Lee, Isc Cohen, Soon Ki Jung, Pricle Filer wih Anlyicl Inference for Humn Body Trcking, IEEE Workshop on Moion nd Video Compuing, 00. [] C. Bregler, J. Mlik, "Trcking people wih wiss nd exponenil mps," CVPR 998, pp.8-5. [3] K. Choo, nd D.J. Flee, People rcking wih hybrid Mone Crlo. ICCV 00, vol II, pp. 3-38. [4] Q. Delmrre, O. Fugers, 3D riculed models nd muli-view rcking wih physicl forces, CVIU, vol 8, no. 3, Mrch 00, pp. 38-357. [5] J. Deuscher, A. Blke, B. Norh, nd B. Bscle, Trcking hrough singulriies nd disconinuiies by rndom smpling. ICCV 999, vol., 44-49. [6] J. Deuscher, A. Blke, I. Reid, Ariculed Body Moion Cpure by Anneled Pricle Filering, CVPR 000. vol, 6-33. [7] M. Isrd, A. Blke, CONDENSATION condiionl densiy propgion for visul rcking, IJCV, 998. [8] M. Isrd nd A. Blke, ICondension: Unifying low-level nd high-level rcking in sochsic frmework, ECCV 998. vol., pp 893-908. [9] I. A. Kkdiris, D. Mexs, "3D humn body model cquisiion from muliple views," ICCV 995, pp. 68-63. [0] J. McCormick nd M. Isrd, Priioned smpling, riculed objecs, nd inerfce-quliy hnd rcking, ECCV 000, vol, pp. 3-9. [] R. vn der Merwe, A. Douce, N. de Freis nd E. Wn, "The Unscened Pricle Filer", in Advnces in Neurl Informion Processing Sysems (NIPS3, MIT Press, Eds. T. K. Leen, T. G. Dieerich nd V. Tresp, Dec, 000. [] I. Mikic, M. Trivedi, E. Huner, P. Cosmn, "Ariculed body posure esimion from mulicmer voxel d," CVPR 00, Vol., pp. 455-460. [3] T. B. Moeslund, Erik Grnum, A survey of compuer vision-bsed humn moion cpure, CVIU, 00, 8(3, pp. 3-68. [4] A. Mohn, C. Ppgeorgiou, T. Poggio, "Exmplebsed objec deecion in imges by componens," PAMI, Vol. 3 Issue: 4, April 00, pp. 349 36. [5] H. Moon, R. Chellpp, A. Rosenfeld, "Trcking of humn civiies using shpe-encoded pricle propgion," ICIP 00, vol, pp. 357-360. [6] R. Plnkers, P. Fu, "Ariculed sof objecs for video-bsed body modeling,", ICCV, 00, vol, pp. 394-40. [7] Y. Rui, Y. Chen, Beer Proposl Disribuions: Objec rcking using unscened pricle filer, CVPR 00, vol., pp. 786-793. [8] H. Sidenbldh, Michel J. Blck, D. J. Flee, Sochsic Trcking of 3D Humn Figures Using D Imge Moion ECCV 000, pp. 70-78. [9] H. Sidenbldh, M. J. Blck, nd L. Sigl, Implici probbilisic models of humn moion for synhesis nd rcking, ECCV 00. vol., pp. 784-800. [0] C. Sminchisescu, B. Triggs, Covrince Scled Smpling for Monoculr 3D Body Trcking, CVPR, 00, vol, pp. 447-454. [] J. Sullivn, J. Rischer, Guiding rndom pricles by deerminisic serch ICCV 00, vol., pp 33-330. [] C. R. Wren, A. Azrbyejni, T. Drrell, A.P. Penlnd, "Pfinder: rel-ime rcking of he humn body," PAMI vol. 9 Issue: 7, July 997 pp. 780-785. [3] T. Zho, R. Nevi, Sochsic Humn Segmenion from Sic Cmer, IEEE Workshop on Moion nd Video Compuing 00. [4] Douce, A., de Freis, J. F. G., Murphy, K., nd Russell, S. (000. "Ro Blckwellised pricle filering for dynmic Byesin neworks," Unceriny in Arificil Inelligence.