Multi-Scale Visual Tracking by Sequential Belief Propagation

Abstract

A novel statistical method is proposed in this paper to overcome abrupt motion for robust visual tracking. Existing tracking methods that are based on the small motion assumption are vulnerable to abrupt motion, which may be induced by various factors, such as the unexpected dynamics changes of the target, frame dropping and camera motion, etc. Although with computational benefits, methods based on hierarchical search is inadequate to this problem because the propagation of the searching error may end up with bad estimates in fine scales. Since different scales contain different salient image features, we propose a new formulation in which searching and matching will be done collaboratively in different scales. The theoretical foundation of this new approach is based on dynamic Markov networks, where the bi-directional propagation of the belief of the target's posteriors on different scales reveals the collaboration among them. A nonparametric sequential belief propagation algorithm for the dynamic Markov network is developed by implementing the collaboration of a set of particle filters. Extensive experiments have demonstrated the effectiveness and efficiency of the proposed method to cope with various types of abrupt motions.