Online Multi-Object Tracking Based On Global Feature Enhancement Network And Historical Feature Complement Iteration
Zunwang Ke, Ping An, Gang Wang, Yugui Zhang · 2024
Multi-object tracking is an important part of computer vision and plays a crucial role in many real-world application scenarios. The existing integrated tracking framework loses the key features of the target when fusing features and iteratively updating the trajectory apparent features, resulting in tracking instability. In order to solve the above problems, this paper proposes an online multi-object tracking method based on global feature enhancement network and historical feature complement iteration. In this method, a dual global feature fusion unit is used to replace the global feature enhancement network of summation or splicing operation, which is used to extract more discriminative key features and enhance the recognition ability of objects in multiple scenes. In addition, the historical features are used to supplement the iterative mechanism, which introduces random features and average features from the historical appearance features stored in the trajectory as supplementary features to supplement the lack of object features in the current frame. Comprehensive experiments are carried out on MOT15, MOT16, MOT17, and MOT20 datasets, in which the MOTA score of 73.2% is achieved on MOT16 dataset, indicating that the proposed method has good tracking performance. Source code: http://github.com/Anbping/gfen-hfci