Multi-template-based Object Tracking via Occlusion Detection

Although object tracking methods are developing rapidly recently, their robustness to occlusion scenarios remains a fundamental problem. In this paper, we focus on single object tracking based on the widely-adopted Siamese network. In order to reduce the frames of lost target, we propose a multi-template based object tracking framework by detecting occlusions. Firstly, a multi-template design is introduced, which stores long-term target appearance in libraries, and updates current template by fusing the similarity-oriented short-term knowledge. Secondly, occlusion detected is applied to guide template updating. Both occlusion and template memories contribute to the insensitivity to the learning rate. We conduct extensive experiments on VOT2018 and VOT2019. The significant reduction of frames of lost target validates the overall effectiveness of our approach.