Abstract

Visual trackan,g cou,ld be treeted as a param,eter estim.ation. problem, of target representastion based on observations in im,age sequ.ences. A rich,er target represen,tation, would i ncw better ch,ances of successfd tracking in cluttered and dyn.a.m.ic en,viron,m,ents. However, th,e dim,ensionmlity of target's state space also increase.s m,&inlg tracking a formlidable estim,ation problem. In this paper; the problem, of trncking and integmtin,g m,vltied in, n probabilistic fra,m,ework and torized graphical m,odel. Stmxtu.red of such gmphicnl m,odel factorizes diflerent m,odalities nn,d suggests 0. co-inference process am,on,g these m,odalities. A sequ.en,tial Mon,te Carlo a,lgorith,m. is propo.sed to give nn. eficien,t o,pproxim.ation of the co-inference based on the hportance sam,plin,g tech,n,ique. This nlgorith)m, is im,plem,ented in, real-ti,m,e at arou.nd 30Hz. Spec(fically, tro,ckin,g both position,,.sh,n,pe and color distribu.tion of a tnrget is investigated in, th,i.s paper. Our exten,sive experim>en,ts show th,at the proposed nlgori,thlmj performs robu,stly in n la1-9e variety of trackin,g scenmrios. The qproach presen,ted in th,is pctper ha,s th,e poten,tial to solve oth,er.sensor fmion, problem,s. 1

Citations