A Survey of Single Object Tracking

This paper aims to comprehensively and deeply sort out the development context and research status of single target tracking technology. The beginning points out the importance of single target tracking in the field of computer vision and its wide application in many practical scenarios, such as traffic monitoring, autonomous driving and mobile robots, and emphasizes the many challenges faced by this task, including background interference, target occlusion, scale and posture changes, and complex motion patterns. Then, in chronological order, the traditional tracking methods (probability-based and feature-based) and deep learningbased tracking methods (divided into four categories: CNN-based, twin network, detection-driven and online optimization, and Transformer) are described in detail, and the principles, advantages and disadvantages of each method are analyzed, and the key network architecture is displayed with clear charts. In the performance evaluation section, the commonly used performance indicators in the field of target tracking (IoU, SR, AUC, P, Average IoU) and four major mainstream data sets (OTB, LaSOT, GOT10k, TrackingNet) are introduced. By comparing the performance results of different methods on these data sets, the evolution of technology and the process of performance improvement are intuitively presented. At the end of the article, the current problems in single target tracking that need to be solved are discussed in depth, such as appearance change and occlusion, multi-target interference and complex background, and immature evaluation system. Possible research directions in the future are also prospected, providing a systematic and practical reference guide for those engaged in related research.