A long-term visual tracking algorithm based on tracking-evaluation-learning-detection

Long-term visual objec tracking is an open and challenging research topic. There is still a large gap between the existing research and the actual application requirements. This paper studies the long-term tracking algorithm, draws on the ideas of the TLD framework, and pays more attention to the role of confidence evaluation in long-term tracking. Taking confidence evaluation as an important part of long-term tracking, this paper proposes a tracking-evaluation-based -Learn-detect long-term tracking algorithm. The main work of this paper can be summarized into three aspects: First, an improved fully-correlation filtering re-detection algorithm is proposed. This algorithm can make full use of the output resources of the confidence evaluation model, effectively reduce the calculation amount during re-detection, and improve Detection efficiency. Secondly, the interaction and update strategies of the short-term tracker, confidence evaluation model, and re-detector in the framework are designed. The proposed strategy can effectively reduce the number of re-detections. Combined with the improvement of the short-term tracking model in the previous article, the reset efficiency of the tracker can be effectively improved after the target is lost. Finally, the validity of the framework is tested on the long-term tracking evaluation dataset.