Multi-camera Tracklet association and fusion using ensemble of visual andgeometric cues

International audienceData association and fusion is pivot for object trackingin multi-camera network. We present a novel frameworkfor solving online multi-object tracking in partially overlappingmulti-camera network by modelling tracklet associationas combinatorial optimization problem hypothesizedon ensemble of cues such as appearance, motion and geometryinformation. Our method learns discriminant weightas a measure of consistency and discriminancy of featurepatterns to make ensemble feature selection and combinationbetween local and global tracking information. Ourapproach contributes uniquely in the way tracklet selection,association and fusion is done. Once multi-view correspondencesare established using planar homography, DynamicTime Warping algorithm is used to make tracklet selectionfor which similarity has to be calculated i.e overlappingtracklets and subtracklets. Then trajectory similarities arecomputed for these selective tracklets and subtracklets usingensemble of appearance and motion cues weighted byonline learnt discriminative function. Later on, we tacklethe association problem by building a k-partite graph andassociation rules to match all the pair-wise trackets. Finally,from outcome of hungarian algorithm, the associatedtrajectories are later fused. Fusion is done based on calculatedindividual tracklet reliability criteria. Experimentalresults demonstrate our system achieve performance thatsignificantly improve the state of the art on PETS 2009