Object Tracking Algorithm based on Transformer with Temporal Contexts
Na Li, Wenhan Jiang, Mengqiao Liu, Wanyv Xin · 2023
Most existing trackers rely solely on current frame information for object tracking. As a result, phenomena such as drifting or tracking failures occur when complex situations arise, such as out-of-view, fast motion and motion blur. To address these issues, this paper proposes a object tracking algorithm based on Transformer with temporal contexts. Firstly, this paper presents a temporal adaptive network, where dynamically calibrated weights are utilized to fuse the feature information from adjacent frames. Specifically, the calibration weights are extracted using the bottleneck. Subsequently, feature enhancement and fusion are performed using a transformer module, which enhances the feature representation. Lastly, on the basis of intersection-over-union loss, an improved localization loss function was designed by incorporating two additional factors, center point distance and aspect ratio, thereby further enhancing the accuracy of object scale estimation. Extensive experiments are conducted on the UAV123 and OTB100 datasets. And compared to four other algorithms, our proposed algorithm achieved the overall best performance. The experimental results show that the proposed algorithm effectively handles complex factors such as out-of-view, fast motion, and motion blur, thereby improving the quality of tracking.