A jump-diffusion particle filter for tracking grouped and fragmented objects

We present a jump-diffusion particle filter for tracking grouped and fragmented supra-threshold targets in visual data. The approach deals with multiple interacting targets in which single objects may yield multiple measurements (target fragmentation), and also for situations in which targets merge and move together. Data association, spatial grouping of features and spatial fragmentation are incorporated as states of the system, which therefore operates in a hybrid (discrete and continuous) space. We outline a method for sorting the particles into subsets corresponding to discrete states, which practically allows a more efficient exploration of the true states of one or more objects at the point of inference in the PF. Evaluation of tracking of both objects and interaction states is demonstrated with several real video sequences that are demonstrably difficult to track.