A Robust Multiple Object Tracking Algorithm under Highly Occlusion

We describe a real-time multiple face-tracking algorithm under highly occlusion. In order to resolve the occlusion and temporal lost problem, a robust data association + filtering procedure is proposed. The mechanism combines the census transform based block-by-block strategy to infer the occlusion state via concerning observation changes of two faces. And a robust and straightforward filtering approach is provided to infer the state of the occluded object, so that the prior motion cues and observations are jointly utilized. Finally cross validation scheme is introduced to adjust the association process and resist unpredicted motion changes. Combining the proposed scheme to resist occlusion with a baseline discriminative kernel tracker, experiments demonstrate that the proposed tracking algorithm has favorable capability on video sequences.