Multiple Target Tracking Using Hierarchical Data Association Based on Network Flows

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%.