ParallelMOT: Pay More Attention in Tracking

Abstract

Modern multiple object tracking has made great progress of the JDE model. Because the JDE model uses a shared model, its calculation speed and accuracy have been greatly improved. But using the same network to predict detection and re- ID will affect each other when the network feedback, thereby reducing the MOTA (Evaluation Measures for MOTChallenge) accuracy, and when the network detects the object and ID information separately, it will greatly increase the computing time. We propose a new MOT method named ParallelMOT, which uses two different branches to reduce the mutual influence of network feedback, and uses object information fusion to improve the feature extraction of the object, and uses a new network model to predict embedding for achieving better MOT accuracy.