Multi-target tracking by detection
Zeng Qiaoling, Gongjian Wen, Dongdong Li · 2016
Aiming at the problem how to express relevant relationship between multiple targets, we propose an approach based on the tracking-by-detection (TBD) strategy, where detections from the HOG classifier are regarded as image evidence. Focusing on the issue of localization uncertainty, data association based on greedy heuristics is executed iteratively to retrieve from the erroneous candidate locations. Additionally, in order to compensate for the deficiencies of greedy heuristics, we put forward a long-interval linking method to correct the per-frame evidence by the aid of Bayesian estimation.