STG-Siam: siamese object tracking based on sparse self attention and graph information interaction
Yining Yang, Jinsong Leng, Hua Cai, Wancheng Liu, Yuhai Li · 2023
Aiming at the problems of insufficient foreground and background discrimination ability of the tracking algorithm in complex scenes, and limited ability to cope with the situation of partial occlusion of the target, we propose a target tracking algorithm based on sparse self-attention interaction with graph information (STG-Siam). To deal with background interference during tracking, a sparse self-attention mechanism is designed to make the model focus more on foreground information and suppress attention to background information. To better handle partial occlusions during tracking, template features and search region features are mapped to graph structures, and local information interaction between graph nodes is used to generate higher quality feature response maps. Experiments on the publicly available datasets VOT2018 and UAV123 demonstrate the effectiveness of the proposed STG-Siam, which can achieve accurate and robust tracking in complex scenarios.