Probabilistic Tracking and Reconstruction of 3D Human Motion in Monocular Video Sequences

Thetrackingandreconstructionof articulatedhumanmotionin 3D is a problemthat hasattracteda greatdealof interestin thelastyears.A systemthatrecovers3D body posefrom videosequences hasapplicationsin vision-basedhuman-computer interaction, marker-lessmotion capture,animation,surveillanceandentertainment suchas computergames. Thefast,non-linearmotionandcomplicatedappearanceof humansandthelarge numberdegreesof freedomof thehumanbodymake thetrackingproblema difficult one. To addresstheseproblems,a systemfor trackingandreconstructionof human motion in 3D shouldpossessthe following: A strongmodel for the appearanceof humansin images;amodelof how peoplemove;andaneffectivestrategy for searching for the right posein eachtime step. In previously presentedsystems,the most commonway of addressingtheseissueshasbeento constrainthe problemdomain. The appearanceof humanscould be constrainedby assumingcertainclothing anda largecontrastbetweenthehumanandthebackground.Furthermore,by addingmore cameraviews,moreinformationaboutthe3D poseof thehumancanbeextractedand ambiguitiesreduced,thusmakingtheproblemeasier . The goal of this thesisis to investigateto which extent the general problemof tracking and reconstructinghumanmotion can be solved, using only a monocular cameraview. Thus, no assumptionsof the appearanceof either the humanor the backgroundareintroduced. Thethesismakesthreecontributions:A probabilisticframework for thearticulated trackingof humanfiguresin 3D; afilter-basedlearnedmodelof humanappearancein imagesandimagesequences; andthreedifferenttypesof modelsof humanmotion, intendedto constrainthesearchin eachtimestepof thetracking.Successful tracking resultsusingthehumanappearancemodelandall threemotionmodelsarepresented. Amongthequestionsleft openis theissueof initialization, a difficult problemin the high-dimensional searchspaceof anarticulatedmodelin 3D. Thecontributionsof this thesisprovideasmallsteponthewaytowardsrobustand accuratearticulated3D trackingof humansin monocularsequences.