Robust Local Effective Matching Model for Multi-target Tracking

摘要

Occlusion is one of the main challenges in multi-target tracking, which causes fragments in tracking. In order to handle with fragments, various motion models were proposed. However, motion model has limited effect on dealing with long-term fragments, because the predictability of target motion declines with increase in fragment length. Thus we propose a robust local effective matching model for partial detections to reduce fragment length first. The proposed model is integrated into a network flow based hierarchical framework to solve long-term fragments step-by-step. Initial tracklets are generated for later analysis in the first level. The robust local effective matching model is used in the second level to reduce fragment length. A motion model is utilized in the third level to solve fragments between tracklets. The benchmark results on 2D MOT 2015 dataset were compared with several state-of-the-art trackers and our method got competitive results with those trackers.