LAUSR

Visual tracking in condition of occlusion has been a challenging task over years. Recently, part-based algorithms have made great progress in handling occlusion. However, the existing part-based methods neglect different importance between central parts and marginal parts. Besides, scale variation remains a difficulty for part-based tracking. In this paper, we propose a novel part-based tracker to solve the above problems. Specifically, we introduce a visual attention mechanism recurrently exploiting co-saliency of target to guide the sampling of parts, which aims to highlight the importance of salient parts and guarantee the semantic integrity so as to improve the robustness handling occlusion. Considering the drift of prediction caused by mutual influence of parts, we implement the non-maximum suppression operation to reduce the high overlaps between parts, and introduce an effective correlation filter as base tracker. To balance the global distribution and local partiality of parts, appropriate update strategy including scale estimation method inspired by particle filters and correlation filters, Hough-voting scheme for target’s center prediction, and principles of part resampling are also fused into the algorithm. The experimental results on VOT 2017 and OTB-50 benchmarks showed that the proposed method is in comparison to the state-of-the-art trackers and good at dealing with occlusion situations particularly.