Robust long-term target tracking algorithm based on objectness-guided tracking

Compared with short-term target tracking, long-term target tracking requires processing of long video sequences and scenes in which the target disappears and reappears. These scenarios are closer to reality, so the research of long-term target tracking algorithms is more practical. In this paper, the long-term target tracking algorithm is divided into two parts, local tracking and global detection. The local tracking module adopts Siam RPN framework, which can not only give the optimal candidate when the target exists, but also accurately determine the target does not exist when the target disappears and enter the global detection module. For the global detection, we propose a new global detection module, which directly predicts the size and position of the target on the whole image based on the target guidance, solves the problem of the original block technology, and greatly improves the tracking speed and accuracy. A large number of experimental results on the VOT-2018LT benchmark show that the proposed method achieves the best performance.