Online Multi-Object Tracking with Three-Stage Data Association
Shi-Jun He, Ping Wang · 2019
Tracking-by-detection is the main paradigm in the field of multi-object tracking. Tracking may fail easily when objects are occluded. In this paper, we propose a method to solve this problem. We adopt a novel three-stage data association. In the first stage, the matching is directly performed according to optimal overlapping if targets and trajectories of overlapping area are very large; in the second stage, Siamese Network is used for the similarity calculation for targets and trajectories with smaller overlapping area and then the Hungarian algorithm is used for matching; the third stage, the remaining trajectories and targets are matched by using the Hungarian algorithm based on the overlapping area. Experiments show that our method effectively improves the tracking success rate and speed.