Multiple Target Tracking with Motion Priors

This paper presents a particle filter-based approach for multiple target tracking in video streams in single static cameras settings. We aim in particular to manage mid-dense crowds situations, where, although tracking is possible, it is made complicated by the presence of frequent occlusions among targets and with scene clutter. Moreover, the appearance of targets is sometimes very similar, which makes standard trackers often switch their target identity. Our contribution is two-fold: (1) we first propose an estimation scheme for motion priors in the camera field of view, that integrates sparse optical flow data and regularizes the corresponding discrete distribution fields on velocity directions and magnitudes; (2) we use these motion priors in a hybrid motion model for a particle filter tracker. Through several results on video-surveillance datasets, we show the pertinence of this approach.