Multi-Object Tracking Based on Feature Fusion and Hierarchical Data Association

Multi-object tracking has been a fundamental topic in recent years. However, most methods can hardly have good performance due to the high appearance similarity, large number of the target objects and error detections in complex scenes. To solve these problems, we propose a multi-object tracking method based on multi-feature fusion and hierarchical data association. Specifically, it uses multi-feature fusion method to improve the anti-jamming ability of data association method and solve the difficulty of tracking accuracy decline caused by large number of objects and error detections. In order to improve the performance of the association method, we refined into multi-layer calculation to solve the problems of objects high similarity appearance and the failure of object trajectory assignment when occlusion occurs. Benefit from our proposed method, on the MOT16 test set, the accuracy of the multi-object tracking method is 44.9%. Compared with the method of multi-feature fusion only, the tracking accuracy of our hierarchical data association is improved by 10.7%.