TLD tracking algorithm for metro construction site security
Jiang Bo Zhu, Yang Chen, Jiaji Cai, Xiujuan Zheng, Huaiyu Wu · 2020
In view of security issues, this paper studies how to track the target at the metro construction site accurately and effectively. It is proposed to replace the tracking module in the TLD(Tracking-Leaning-Detection) algorithm with the GOTURN algorithm, to enhance the tracking robustness of the original algorithm in response to the target's illumination changes, rotation, occlusion, etc. To solve the problem of inaccurate sample division in the TLD algorithm learning module, a sample constraint mechanism is introduced to divide positive and negative samples more objectively and accurately. The sample deletion mechanism of weighted model matching is introduced to delete samples that are weak in characterizing the current target, which improves the tracking accuracy and efficiency of the algorithm. The experimental results show that the improved algorithm can well reduce the interference that is caused by the target pedestrian's illumination change, turning, occlusion, etc. It can accurately track the target pedestrian for a long time, and has good robustness and real-time performance.