Visual tracking based on multi-cue framework and hierarchical aggregation
Yi Zhang, Guixi Liu, Hanlin Huang, Ruke Xiong · 2021 International Conference on Control, Automation and Information Sciences (ICCAIS) · 2021
Existing Siamese network based trackers define target tracking as a similarity matching problem between the initial template and search regions. However, the initial template contains extremely limited target information that ignores appearance variations over time. Aiming at this problem, a visual tracking method based on multi-cue framework and hierarchical aggregation is proposed in this paper. Firstly, we put forward a stability evaluation mechanism to select the stable template for similarity matching, so that the tracker can adapt to the target appearance changes. Secondly, we construct a multi-cue tracking framework combining the initial template branch and the stable template branch for parallel tracking, thus solving the single template limitation. Finally, we propose a hierarchical aggregation strategy that combines complementary merits to further improve tracking performance. Comprehensive experimental results on the benchmark dataset demonstrate that the proposed algorithm achieves favorable tracking performance.