Obect Modelg for Mltcamera Correspodece Usg Fzzy Rego Color Adacecy Graphs Amr Hosse Khall 1 ad Shohreh Kasae 2 1 Sharf Uersty of Techology, Tehra, Ira a_khall@ce.sharf.ed 2 Sharf Uersty of Techology, Tehra, Ira skasae@sharf.ed Abstract I ths paper, a oel mog obect modelg stable for mltcamera correspodece s trodced. Takg to cosderato the color ad moto featres of foregrod obects each depedet deo stream, or method segmets the exstg mog obects ad costrcts a graph-based strctre to mata the relatoal formato of each segmet. Usg sch graph strctres redces or correspodece problem to a sbgraph optmal somorphsm problem. The proposed method s robst agast aros resoltos ad oretatos of obects at each ew. Or system ses the fzzy logc to employ a hma-lke color percepto ts decso makg stage order to hadle color costacy whch s a commo problem mltew systems. The comptatoal cost of the proposed method s made low to be appled real-tme applcatos. Also, t ca sole the partal occlso problem more precsely tha the Meashf occlso soler by 15.7%. Keywords Obect Correspodece, Mltcamera Trackg, Rego Adacecy Graph, Fzzy Color Modelg. 1. Itrodcto The creasg demad for aalyzg mog obects behaor wde area has heghteed the eed for robst modelg ad correspodece of the mog obects scee. The prpose of obect modelg s takg ot some featres from mages of a obect so that the selected featres are stable ad relable ad cold exactly dscrmate the target coseqece proectos. Ths task s dffclt becase a obect motorg systems ca moe fast ad predctably, ca appear a arety of poses ad colors, ad are ofte srroded by cltter. For applcatos sch as otdoor deo srellace, where the etre scee s ot coerable wth a sgle camera, a dstrbted camera etwork shold be mplemeted. I mltew framework, dffereces camera drectos, dstaces to target, llmato statos ad qalty of captrg case sgle obect s looked dfferetly each feld of ew. Chages sze, ew drecto, lmace ad color ales of obects are some artfacts whch make the problem more challegg. For that reaso, specal obect model s eeded to mata correspodece of two-dmesoal proectos of a obect see dfferet ews. Wde arety of methods has bee reported to model mog obects mltew applcatos, bt a few of them hadle color arato ad occlsos effectely wthot employg hgh-leel reasog procedre ad predefed model of target obects. I or experece the low leel mage featres play a crcal role. The proposed method for obect modelg ams to segmet mog obect effcetly accordg to ts color ad moto featres ad accommodate hma-lke color percepto to deal wth certaty obsered colors. The proposed method costrcts a rego adacecy graph for each mog obect perspecte to book relatoal state of each segmet ad tres to redce mog obect correspodece problem to a sbgraph optmal somorphsm problem. The proposed method does t eed ay hgh-leel defto of mog obects ad cold deal wth partal occlsos effcetly. The rest of paper s orgazed as follows. I secto 2 a oerew o related preos works s ge. I secto 3 dfferet steps of the proposed algorthm are trodced detal. The expermetal reslts are show secto 4. Fally, secto 5 cocldes the paper.
2. Preos Work Costrctg obect model has bee extesely stded mage retreal applcatos where the prcple s extractg stable featres for target obect ad lookg for them set of mages. Most sccessfl dea ths area employ tme cosmg algorthms sg waelet, Gabor flter or Hogh trasform [1,2,3], whch are ot applcable real-tme applcatos sch as eromet motorg. I cotext of trackg, obects are modeled dfferetly. Drg trackg wth sgle camera, obects cold be modeled accordg to ther temporal stats formato cldg posto, temporal elocty, appearace featres ad color propertes. These kd of modelg s robst ad stable eromets clded people ad ehcles [4]. Whle trackg mlt cameras etwork, selected featres shold hae lttle chages from dfferet ews, so that specfc obect cold be dscrmate easly from other cameras to track. Varato cameras sed a etwork ad ther dfferet postos, drectos ad dstaces to target, addto to dfferet llmato codtos case selecto of stable featres dfferet ews to be a challegg problem. Most mltcamera applcatos se color formato of target obects as modelg parameters. [5,6] sed aerage color of mog obects HSV space. [7,8] se oly oe color hstogram model per obect. Althogh color hstograms are rotato/scale arat, comptatoally effcet ad robst to partal occlso, bt preet dfferetatg a perso wearg ble eas wth a red shrt from a perso wearg red pats ad a ble shrt. Also, ordary hstogram matchg approaches fal to dscoer qeess of sesed colors mltcamera applcato where a sgle color sesed dfferet ale by commo color spaces. Referece [9] tred to fd ad correct sch dstorto throgh a trag step. Referece. [10] sed Msell color space [11] for costrctg hstograms. I ths approach colors are coarsely qatzed to 11 predefed bs. [12] sed color-spatal dstrbto of mog obect by parttog mog blob ts polar represetato. Some mltcamera perso trackg methods partto body to predefe mber of segmets ad track color formato of each depedetly. Referece [13] clams that color formato of body, pels, ad feet are the most stable formato ee dstrbted cameras. I [14] athors clster a perso accordg to ts color ad moto featres sg watershed ad K-meas algorthm to recogze gestres ad track each. Oe approach to the detecto ad trackg problem s to ft explct obect models of shape, sch as rgd wreframe CAD models [15,16] or flexble acte shape models [17]. Some model fttg approaches focsed o hgh-leel reasog [18,19]. They predefe a model for ther target obects, ad try to ft obects proectos to t ad estmate obect s pose other ews. The robstess of sch approaches s hghly depeds o the defed model. Models coer lmted rage of obects ad behaors ad fals f the mog obects moe fast or do t obey the defte behaors. For sch methods calbrato data of cameras shold be kow. Occlsos, low resolto mages ad arety of poses ad colors whch are commo most applcatos are some crcal perl for robstess of these methods. 3. Proposed Method The sbseqet steps of the proposed mog obect modelg for mltew correspodece are explaed detal the followg sbsectos. 3.1. Mog Obect Segmetato Backgrod sbtracto method [20] sed for extractg foregrod regos each camera. Foregrod regos are the modeled as rego adacecy graph ad are tracked depedetly each frame of each ew sg method defed [21]. Needg to be real-tme ad obect depedet, or segmetato method dffers to what sed [14]. I or mplemetato two featres sed to segmetato foregrod obects: color formato ad drecto of moto ectors. I ths regard we sed Hor & Schck algorthm [22] to compte optcal flows ad Cr ad Cb chaels of YCrCb color space as color featre. The adatage of sg moto ectors segmetato processes s that the drecto ad dstrbto of the optcal flow ca be sed to dstgsh the dfferet mog parts of obects. Ee thogh these parts may hae smlar color oe ew, they sally dffer others. By segmetg homogeos color parts accordg to ther moto, o extra process s eeded to splt ad merge dfferet part of obect model to match them dfferet ews. I addto, moto s sally a relable featre to detect hma actty door eromet ad from dfferet otlooks [14]. Before segmetato a preprocessg s eeded to smooth solated pxels whch ther color or optcal flow do t cofrm wth ther eghbors. We sggest sg blateral smoothg flter whch s preseted (1). h ( x) = f ( ξ ) C( ξ X ) S( f ( ξ ) f ( x)) dξ (1) I whch both the closeess fcto, C, ad the smlarty fcto, S, are Gassa fctos of the Ecldea dstace betwee ther argmets. It s a smple, o-terate scheme for edge-preserg smoothg. It replaces the pxel ale at x wth a aerage of smlar ad earby pxel ales ad does t trodce false colors arod the bodares. By mplemetg sch flter 3-5 tmes, we obta graphc lke appearace whch fe textre has goe. The desred segmetato algorthm mst hae low complexty ad strog power of dscrmato. The choces ths regard are qte ast. We adopt the approach trodced [21] becase of three maor reasos: 1) that method t has bee tred to act po global characterstcs of pxels. 2) It has low
comptatoal complexty ad ca be applcable early lear tme. 3) The graph-based strctre of ths algorthm hghly coforms to or rego adacecy graph (RAG) model. I ths sese root odes of segmetato algorthm whch represet dfferet regos correspod to odes of RAGs. Two odes RAG are adacet f ther correspodg odes segmetato algorthm are adacet more tha predefed threshold pxels. Fg. 1 shows a artfcal mage ad ts RAG yelded by color segmetato. Fg. 2 shows reslts of or proposed foregrod segmetato method a atral scee ad ts costrcted RAGs. Whe there s o hgh-leel descrpto of mog obects, RAG modelg ges sefl strctre to descrbe mog obect ad ts parts. For the prpose of correspodece dfferet ews some relable ad stable featres of regos shold be assocated to each (a) (c) (d) Fg. 1. a) A sample sythetc mage. b) Rego adacecy graph prodced from color segmetato sg [21]. c) He hstogram of Rego 1. d) He hstogram of Rego 3. (a) (d) Fg. 2. Reslt of proposed segmetato a atral scee. a) Orgal corrdor frame. b) Foregrod segmetato of a. c) Rego adacecy graph of b. d) Orgal frame of frot ew. e) Foregrod segmetato of d. f) Rego adacecy graph of e. (b) (e) (b) (c) (f) ode of RAG. Color of regos s sally sed as a robst featre sch applcatos. Howeer, commo color spaces do ot preset relable ad stable featre. I the ext secto or hma percepto-based color modelg s trodced to sole ths shortcomg. 3.2. Hma Percepto-Based Color Modelg Color featres are oe of aalable featres mltcamera trackg applcatos. Dfferece ew drectos, target dstaces, llmato statos ad captrg qaltes cases sgle color of obect s sesed dssmlar each ew to the exted that hma faces a dlemma to recogze the obect. Usg commo color spaces, the ales whch are assged to the sesed colors dffer sch that commo smlarty measremet fctos fod them lke ad msmatch them correctly. I or experece, mltew applcatos hma classfes colors to a lmted set whch s more stable dfferet ews. Mappg colors of obect proecto a ew to members of ths set, hma looks for ay arragemet of colors other ews that cofrms wth hs target model. Hma color classfcato s ot corse. He some tmes faces a dlemma abot a color ad caot determe to whch member the specfc color exactly belogs. I sch a stato hma assgs a membershp ale to each member to whch the color s smlar. Corse parttog the color space to a relately small mber of corse segmets sch as [11] does ot prode a optmal solto for correspodece problem. To help or system to hae hmaod color classfcato, we sed HSL color space ad fzzy logc systems, ad prode fast color segmetato sg lgstc rles of hma tto. HSL s a color space whch s more tte ad closer to the hma percepto of color. Fzzy logc s sed as a terface betwee logc ad hma percepto. Sce HSL color space each color s defed by three ales (H,S ad L), the fzzy logc model has three atecedet arable (He, Satrato ad Lmace) ad oe coseqet arable, whch s a lgstc color. The fzzy sets of the atecedet fzzy arable He are defed based o 6 basc hes whch are dstgshable by hma percepto. We ca see sch classfcato detfcato of dfferet colors of rabow. As show Fg. 3.a, the basc hes are Red, Yellow, Gree, Aqa, Ble, Prple. The membershp fcto of each s show Fog. 3.b dstrbted 0-180. Satrato s defed sg three fzzy sets Gray, Medm, Clear, as show Fg. 4. Lmace s also defed sg the three fzzy sets Dark, Medm ad Brght as descrbed Fg. 5. The coseqet part of each fzzy rle s a crsp dscrete ale of the set Black, Whte, Red, Orage, Yellow, Gray, Brow, Aqa, Ble, Gree, ad Prple. The members of ths set are called Lgstc colors.
Medm (a) Fg. 4. Membershp fctos of satrato pt arable. Membershp fctos hae a great oerlap de to the fact that satrato chages extremely betwee ews. (b) Fg. 3. a) HSL color space. 6 basc colors dstgshable wth hma percepto are deoted by the "*" sg. b) Expermetal membershp fctos of basc Hes. By defg set of fzzy rles accordg to hma belef ad obserato, percepto of lgstcs colors based o atecedet arables (He, Satrato ad Lmace) are traslated to system, so that system cold ferece the lgstc colors ad ther membershp belees smlar to the hma. For example, the rle Yellow Medm Brow s defed by maally classfyg the color prodced by the HSL trple sch that the ales of H, S ad V are the pots of maxmm of the membershp fctos assocated wth the fzzy sets Yellow, Medm ad Medm. Based o the membershp fctos descrbed Fg. 3.a, 4 ad 5, ths case the ales are H = 25, S = 128, L =150. As aother example the rle Ble Medm Brght Aqa s the rle whch s actated mostly whe H = 120 S = 125, L = 220. The color prodced by ths HSL trple wold be classfed by most hma obserers as Aqa. Ths correspods to the atral lagage hma percepto-based rle f the he s Ble, the satrato s Medm ad the ale s Brght the the color s Aqa. Ths, the set of fzzy rles cldes rles sch as: Red Gray Brght Whte Red Medm Dark Black Red Gray Dark Black Red Gray Medm Gray Yellow Clear dark Orage Red Clear Medm Red Ble Medm Brght Aqa Aqa Clear Medm Aqa Sce the model has 6 fzzy set for He, 3 for Satrato ad 3 for lmace, the total mber of rles reqred for ths model s 6 3 3 = 54. The reasog procedre s based o a zero-order Takag-Sgeo model. I proposed method, base o H, S, L ales of a pxel membershp belefs to each lgstc color s Fg. 5. Membershp fctos of lmace pt arable. Membershp fctos hae a great oerlap de to the fact that satrato chages extremely betwee ews. compted. Accmlate belefs of rego s pxels to each 11 lgstc colors, costrct a hstogram. After ddg each cell of the hstogram to area of the rego, the t s called lgstc color hstogram. For each ode of RAG a lgstc color hstogram s assocated. 3.3. Matchg Two Fzzy Rego Adacecy Graphs Smlarty betwee two obects modeled term of RAGS ca be approxmated by measrg sbgraph optmal somorphsm of ther graphs. Formally, two graphs are sad to be somorphc, f there s a oe to oe correspodece betwee ther ertces ad edges sch that cdece relatoshp s presered. If the somorphsm s ecotered betwee a graph ad a sbgraph of aother larger graph, the the problem s called sbgraph somorphsm. Whe odes of graphs are attrbted ad perfect matchg betwee attrbted ales of graphs s target the problem s called sbgraph optmal somorphsm. De to chages drecto, sze ad colors of proectos of a obect, we caot expect to hae graph somorphsm betwee ewed RAGs of a sgle obect. Therefore t s more robst to se sbgraph optmal somorphsm as smlarty measremet. Also, thaks to sbgraph optmal somorphsm we ca dscrmate RAGs of dfferet obects case of partal occlso. The ma drawback of sbgraph optmal somorphsm les ts heret comptatoal complexty. The sbgraph somorphsm problem s kow to be NPcomplete, ee for plaar graphs. May efforts hae bee drected to fd effcet algorthm for ths prpose. We approxmate ths smlarty measremet by adaptg method trodced [23]. Ths method ses mmm cost seqece of elemetary graph maplato operators as smlarty measremet betwee two graphs. These operators whch trasform oe graph, say, to the aother graph, say h, cosst of: 1) deletg a ode or
a edge from, 2) sertg a ode or a edge to, 3) sbstttg a ode from by a ode from h. Accordg to [23] t s better to decompose the two caddate RAGs to smaller sbgraphs called Basc Attrbted Relatoal Graphs (BARG). A BARG s oe leel tree, cldg a ode (deoted by ) ad all ts eghbors. Fg. 6 shows dfferet BARGs of preseted RAGs Fg 2.d. (a) (b) (c) (d) (e) (f) (g) (h) () () (k) (l) Fg. 6. (a-h) Basc attrbted graphs of rego adacecy graph of Fg. 2.c. (-l) Basc attrbted graphs of rego adacecy graph of Fg. 2.e. Sbgraph optmal somorphsm measremet ca be calclated wth optmal matchg of a complete weghted bpartte graph whose odes are BARGs of each RAG parttoed two grops. Weght of each coectg edge bpartte graph s smlarty dstace betwee BARGs at two tals of the edge. Smlarty dstace of th BARG of RAG camera C1, to th BARG of RAG camera C2, s compted sg (2). Dst(, ) = Dst( e, ) deg( ) deg( + Dst bpartte ( N( + w deg( + w deg( ) ) ) ), N( whch Dst(, ) deotes smlarty measremet betwee root odes of two BARGs whch s compted sg adapted accmlate fzzy tersecto formla (3). Dst(, ( w ) = x, y y, x lgst colors m(. hst( x), )). hst( x))) 2 (2) (3) c1, x, y w s a coeffcet whch trodces smlarty betwee two lgstc colors two cameras. For example It s lkely to st a gray color oe ew as black oe aother ew. So we defe w gray, black = 0. 5. Smlarty of black ad whte s ery low compare wth smlarty of black ad gray, therefore we hae w whte, black << 0. 5. I other words, w x, y ca be see as the sbstttg cost of two odes. We ca defe these coeffcets maally accordg to color dgresso see two cameras, c1 ad. Also, t s possble to obta the coeffcets accordg to see colors of defte mog obect a trag step. Fdg two regos are correspodece, the w x, y ale of ther maor lgstc colors are creased by α factor. w x, y s are pdated depedetly for each perspecte mage of a obect each two cameras. Whe two regos hae great mber of pxels commo at specfc lgstc color, formla (3) creases ther smlarty measremet more sgfcat tha stats whch they hae same amot of tersecto bt dstrbted more lgstc colors. deg( ) (2) deotes degree of root of th BARG of RAG. w ad w e deote ode ad edge serto costs. I or experece t s better to defe w ad we adapte ad as erse fcto of dfferece sze of foregrod obects. If two caddate foregrod obects are same sze, w ad w are assged to ther maxmm ales. Icreasg sze dfferece of two foregrod obects, cases decrease amot of these coeffcets. I ths maer we ca effcetly hadle dfferet resoltos of obect dfferet cameras. I case of occlso, w ad we do t chage ad ther last amots are sed. Ths, t s ecessary to arche amots of these arables for each BARG of each two caddate RAGs. Dst ( N( ), N( )) s calclated as the bpartte maxmm matchg of a complete weghted bpartte graph costrcted from eghbors of ad. Neghbors of, N ( ), are parttoed oe had ad eghbors of, N ( ), o the other had. The weght of coectg edges s compted sg (3). Fg. 7 shows sch bpartte graph for BARG c ad BARG of Fg.7. Bpartte graph costrcted from eghbors of root of c ad BARGs of Fg.6. Root odes ad all ther coectg edges do ot partcpate bpartte graph ad are oly demostrated for trasparecy. e
Fg. 6. 3.4. Trackg Usg Rego Adacecy Graphs Fg.8 demostrates a oerew of or mltew trackg model. Fg.8 Block dagram of proposed mltew tracker. Usg methods trodced sectos 3.1 ad 3.2, fzzy rego adacecy graph for each foregrod obect of each camera s prodced. For trackg obects each camera we se combato of seqetal Kalma flter ad ot probablstc data assocato (JPDA) [25]. Smlarty measremet trodced secto 3.3 s sed to fd most feasble hypothess aldato gate of JPDA. For trackg sgle camera we set w 1 ff x, y = x = y ad w x, y = 0 f x y. It s de to the fact that we hae ot color dstorto a sgle ew. Best matchg coseqece betwee RAGs of two ews s sed to defe correspodece betwee obects. Fdg two mog regos are correspodece, w x, y of ther maor lgstc colors are pdated, secto 3.3. I case of occlso, RAG of foregrod place of occlso s called Sper Rego Adacecy Graph (SRAG). Tracker cold segmet occldg obects from occlso area by examg sbgraph somorphsm of preos RAG of each partcpatg obects. The RAG whch ca be segmeted completely from SRAG s RAG of most closed obects. By remog foded RAGs from SARG we ca do sbgraph somorphsm process aga to locate other obects. The proecto of obect s bodg box from other ew to cosderg ew, lmts or search throgh SARG. 4. Expermetal Reslts To ealate sefless of or proposed method, we mplemeted the mltobect trackg system C++ sg OpeCV frastrctre. We r or code o CPU 3.00 GHZ wth 2.00 Ggabyte RAM. Fzzy rles are costrcted Matlab eromet ad ts prodced.fs fle s loaded or program. We hae sed otdoor atral scees, whch are cosdered to be more challegg for color segmetato ad correspodece methods. For ths prpose, we sed the CAVIAR[25] ad PETS2001[26] as or datasets. CAVIAR cossts of seqeces across the hallway a shoppg cetre Lsbo.. For each seqece, there are two sychrozed orthogoal deos, oe wth the ew across the hallway ad the other alog t. The format s half-resolto PAL stadard (384 288 pxels, 25 frames per secod). PET2001 cossts of otdoor seqeces take from a ersty locato. Each dataset cldes two deos ewg eromets each of whch captred wth 30 frames per secod ad frame sze 768 520. We reszed each frame by half to crease or processg speed. Accordg to the dstace amog the locato of mog obects ad cameras, the captred deos mostly prode low resolto obects. Ths fact cases a maor problem for most aalable methods. 4.1. Mog Obect Segmetato Fg. 9 shows three coseqet frames of TwoEterShop1 seqece of CAVIAR captred alog a corrdor ew. It s otceable how well dfferet parts of the froter perso are segmeted (head, hads, body, legs ad feet are segmeted effcetly). Howeer, by gog far ad ths decreasg resolto of mog obects (e.g., far persos Fg. 9) the accracy s redced, bt we stll obta meagfl segmets. The abormally segmets the mddle of these frames are de to false backgrod segmetato. Fg.10 shows or reslts obtaed from for par frames of OeStopNoEter2 seqeces CAVIAR. Fg. 2 (a) (b) (c) (d) (e) (f) Fg.9. Three coseqet frames of TwoEterShop1 seqece of CAVIAR alog the corrdor ew. Colors assocated to each segmet each frame are chose radom.
(a) (b) (c) (d) (e) (f) (g) (h) () () (k) (l) (m) () (o) (p) Fg. 10. Sample reslts obtaed for OeStopNoEter2 data set of CAVIAR. (a-b) Orgal frames of corrdor ew. (e-h) Or segmetato reslts for (a-b). (-l) Orgal frame of frot ew. (m-p) Segmetato reslts of (-l). colors segmeted parts are chose radom. preseted Secto 3.1 was obtaed from these seqeces as well. Table 1 lsts the complexty cost of or proposed method for mog obect segmetato ad Fzzy color assgmet processes. Or expermetal reslt showed that or fast ad accrate segmetato algorthm s speror to [14]. Obosly, the comptatoal cost of the methods s hghly depedet to the sze of mog obects. Table 1: Performace comparso of dfferet mog obect segmetato methods. Vdeo Type Frame Sze Aerage elapsed tme for each frame secods. Method Proposed Proposed Method [14] TwoEterShop1 corrdor ew 384 288 1.5167 0.0613 4.2. Trackg Reslts Usg Fzzy Rego Adacecy Graph To ealate the performace of or tracker, plots show Fg. 11 are show. The plot shows the trackg error of selected tracks from the frst PETS2001 deo seqece, ad has bee aotated to dcate the key eets that hae occrred drg the perod of trackg. I Fgs. 11.c the ble ad red colors show the error of occlso soler sg the Meashft ad proposed method, respectely. It ca be obsered that the largest errors occr whe the obect splts to seeral segmets de to occlso. Or proposed method has redced the error by abot 15.7%. CAVIAR PETS 2001 TwoEterShop1 frot ew OeStopNoEter2 corrdor ew OeStopNoEter2 frot ew TestSet1 Camera1 TestSet1 Camera2 384 288 0.0634 0.0500 384 288 1.3231 0.0827 384 288 0.0534 0.0500 384 260 0.3458 0.0603 384 260 0.5765 0.0701 (a) (b) (c) Fg. 11. Reslt of proposed tracker for Pets2001 dataset. a) Before occlso. b) Drg occlso. c) Error of occlso solers betwee frames 888 ad 937: ble: Meashft method, red: proposed method.
5. Coclso I ths paper, we proposed a oel method to model mog obects mltcamera systems. Or proposed method costrcts a rego adacecy graph for each foregrod obect ad soles the correspodece problem throgh a sbgraph optmal somorphsm approach. A Fzzfed color hstogram s the assocated to each graph ode as ts attrbte. Or proposed method s fast eogh to be employed real-tme applcatos. Also, or proposed method s robst to face partal occlso, color chages, ad dfferet obect resoltos ad oretatos. Ackowledgmet Ths work was part spported by a grat from ITRC. Refereces [1] T. Egeo, M. Potl, ad C. Papageorgo, ad T. Poggo, Image represetatos for obect detecto sg kerel classfers, I Asa Cof. o compter so. Pp.687-692, 2000. [2] V. Kyrk, J.K Kamarae, ad H. Käläe, Smple Gabor featre space for arat obect recogto, Patter Recogto Letters, Vol. 25, No. 3, pp. 311-318, 2004. [3] M. Ulrcha, C. Steger, ad A. Bamgarter, Real-tme obect recogto sg a modfed geeralzed Hogh trasform, Patter Recogto, Vol. 36,No. 11, pp.2557-2570, No. 2003. [4] S.A Velast, ad P. Remago, Itellget dstrbted deo srellace system, Lodo: The sttte of Electrcal Egeers, 2005, pp. 164. [5] T. Hag, ad S.T Rssell, Obect Idetfcato a Bayesa Cotext, IJCAI, 1998, pp 1276-1283. [6] V. Kettaker, ad R. Zabh, Bayesa Mlt-Camera Srellace, Cof. o compter so ad patter recogto, pp.2253-2261, J. 1999. [7] J. Orwell, P. Remago, ad G.A. Joes, Mltple Camera Color Trackg, IEEE Itl. Workshop o sal srellace, pp. 1355-1360, Oct, 2003. [8] A. Elegmmal, ad L.S. Das, Probablstc framework for segmetg people der occlso, IEEE Itl. Cof. o Compter Vso,2001. [9] O. Jaed, Z.Rasheed, K. Shafqe, ad M. Shah, Trackg Across Mltple Cameras wth Dsot Vews, 9 th IEEE It. Cof. o compter so, pp. 952-957, Nce, Frace, 2003. [10] A. Glbert, ad R. Bowde, Icremetal Modelg of the Posteror Dstrbto of Obects for Iter ad Itra Camera Trackg, Brtsh mache so coferece, Oxford, UK, 2005. [11] S. Blackma, ad R. Popol, Desg ad aalyss of moder trackg systems, Artech Hose, 1999. [12] J. Kag, I. Cohe, ad G. Medoo, Cotos trackg wth ad across camera streams, IEEE compter socety Cof. o compter so ad patter recogto, ol. 1, pp. 267-272, Je. 2003. [13] W. Zadel, ad B.J Krose, A seqetal Bayesa algorthm for srellace wth o-oerlappg cameras, It. Joral of patter recogto ad artfcal tellgece. Vol. 9, No. 8, pp 977-996, Dec. 2005. [14] H. Aghaa, ad C. W, Layered ad collaborate gestre aalyss mlt camera etworks. I It. Cof. o acostcs, speech ad sgal processg, Vol. 4, No. IV, pp 1377-1380, Hooll, USA, Apr. 2007. [15] M. Frak, H. Haag, H. Kollg, ad H.H. Nagel, ``Trackg of occlded ehcles traffc scees, 4 th Eropea Cof. o compter so, ol. 1065, pp. 485-494, Cambrdge, UK, 1996. [16] D. Koller, K. Dalds, ad H.H Nagel, ``Model-based trackg mooclar mage seqeces of road traffc scees, It. Joral of compter so, Vol. 10, No. 3, pp. 257-281, 1993. [17] A. Bamberg, ad D. Hogg, Learg flexble models from mage seqeces, 3 th Eropea Cof. o compter so, ol. 800, pp. 299-308, Stockholm, Swede, 1994. [18] D. Ramaa, D.A Forsyth, ad A. Zsserma, Trackg people by learg ther appearace, Tras. O patter aalyss ad mache tellgece. Vol. 29, No. 1. Ja 2007. [19] A. Jepso, D. Fleet, ad T. El-Maragh, Robst ole appearace models for sal trackg, IEEE tras o Patter aalyss ad mache tellgece, Vol. 25, No. 10, pp. 1296-1311, Oct. 2003. [20] A. Darab, A.H. Khall, ad S. Kasae, Graph-based segmetato of mog obect, press o It. GCC IEEE, Maama, Kgdom of Bahra, No. 2007. [21] P.F. Felzeswal, ad D.T. Htecheler, "Effcet Graph- Based Image Segmetato," It. Joral of compter so, ol. 32, o. 14, 2004. [22] B.K.P. Hor, ad B.G. Schck, Determg Optcal Flow, Artfcal Itellgece, Vol. 59, pp. 81-87, 1993. [23] Y. El-Sobaty, ad M.A Ismal, A ew algorthm for sbgraph optmal somorphsm, Tras. O patter recogto, Vol. 31, No. 2, pp. 205-218, 1998. [24] Y. Bar-Shalom, ad X. L, Mlttarget-mltsesor trackg: prcples ad techqes, YBS pblshg, 1995. [25] EC Fded CAVIAR proect/ist 2001 37540, fod at URL: http://homepages.f.ed.ac.k/rbf/caviar/. [26] Fod at URL http://pepa.essex.ac.k/pa/px/pets/pets2001/datase T1/.