基于网络流的分层关联多目标跟踪

Multiple Target Tracking Using Hierarchical Data Association Based on Network Flows

为解决在有遮挡、光照变化等复杂环境下的多目标跟踪问题,提出一种基于最小费用流的多目标跟踪算法.该算法吸取分层数据关联的思想,将多目标跟踪分成2个阶段.首先利用双阈值法对检测器输出的响应进行初级关联,形成可靠轨迹片;然后根据轨迹片形成的有向无环图,利用最小费用流法求解进行进一步的数据关联,得到长轨迹.为处理遮挡问题,关联过程中对目标采用分块处理的方式.在公共数据集上的实验结果表明,文中算法能够在复杂场景中有效地实现多目标跟踪,平均跟踪准确度和精度达到85.9%和83.7%,并对遮挡、轨迹分段等问题具有良好的解决效果.

A tracking-by-detection algorithm is proposed based on network flows using hierarchical data as- sociation for multi-target tracking in complex environment. The multi-target tracking problem could be di- vided into two phases using hierarchical data association. Firstly, a dual-threshold strategy was adopted to perform low-level association and generate reliable tracklets between adjacent frames according to detection responses. Then, the minimum cost flow algorithm was used to obtain the long trajectories based on the di- rected acyclic graph formed by tracklets. We adopted part-based appearance model during the process of association considering occlusion in the video. Experiments show that the proposed method can handle oc- clusion and trajectory fragment effectively in complicated situations with average tracking accuracy and precision of 85.9% and 83.7%.