Object tracking based on foreground adaptive bounding box and motion state redetection

Siamese network is successfully applied in object tracking. Most of the existing Siamese tracking methods extract template features in the first frame, which will cause the tracker to ignore the appearance change of the target in the subsequent video. In this paper, we propose a tracker based on foreground adaptive bounding box and motion state redetection. The tracker infers the reliability of tracking by the motion pattern of the bounding box. When an anomaly is detected, the tracker will redetect using the continuously updated template. Furthermore, our tracker employs an adaptive bounding box to avoid the effects of inaccurate rotation of the bounding box. The results on the VOT2018 dataset show that our tracker achieves stronger robustness and higher accuracy, providing superior performance compared to the current state-of-the-art trackers.