A Dual Scale Matching Model for Long-Term Association

Multi-object tracking can be characterized as a data association problem. The advantage of RNN in processing temporal dependence makes it an ideal selection in data association. When factors such as scene congestion and weak illumination cause detection failure especially long intervals, association is often very difficult and eventually leads to tracking failure. To solve this problem, Dual Scale Matching model (DSM) containing a Motion Trend Match Network (MTMNet) and an Appearance History Memory Network (AHMNet) is proposed. DSM is a long-term optimization method based on multiple hypothesis tracking. MTMNet aims to learn a similarity metric matching function between tracklets leveraging the motion feature. AHMNet is designed to provide optimal pruning strategies leveraging long period appearance feature. Our method is effective on MOT17 benchmark and it shows that we achieve considerable competitive results with current state-of-the-art trackers.