Collaborative tracking based on contextual information and local patches

Robust and accurate tracking is a very difficult task in most challenging situations, such as occlusion and background clutter. A collaborative approach based on contextual information and local patches is proposed in this paper. Firstly, we represented the target by using a series of patches, and each patch independently performs a tracking task based on correlation filtering. To identify and utilize reliable patches, we propose a measure by using reliability and stability indices. Secondly, to suppress background noise, we used double bounding boxes to represent the object consists of the background and foreground. Furthermore, a fine search strategy based on the Bayesian inference framework is adopted to enable the proposed tracker can robustly track various appearance changes. To achieve accurate tracking performance in the long-term tracking process, we used the confidence map to distinguish whether the patch is under severe occlusion. The experimental results show that our approach outperforms several existing state-of-the-art algorithms on the challenging benchmark datasheets.