Joint detection and embedding of multi-object tracking with feature decoupling

Multi-object tracking (MOT) represents a core challenge in the field of computer vision and is extensively utilized across various industrial applications. Recently, joint detection and embedding approaches in MOT have gained significant attention due to their distinctive architecture, which enhances model reusability and accelerates inference speed. However, this approach has a limitation: the detection task relies on inter-class features, whereas the re-identification (ReID) task depends on intra-class features. Directly merging two distinct tasks can lead to competition between them, ultimately resulting in a decline in tracking accuracy. In this paper, we improve the optimization conflict problem in the training process of joint and embedded MOT algorithms. Spatial based Decoupling Module (SDM) is proposed to decouple the backbone features in the spatial dimension, which can adjust the spatial attention of each channel of each subtask feature, and use strip convolution to ensure the inference speed. Channel based Decoupling Module (CDM) is proposed to decouple the backbone features in the channel dimension. The effect of the combining methods of SDM and CDM on the performance of the overall MOT model is explored, and the optimal combination is combined to form the Spatial and Channel based Decoupling Module (SCDM) module. The baseline method used in this study is FairMOT, and we use this method to evaluate the performance of SCDM on the MOT16, MOT17, MOT20, and DanceTrack datasets. Experimental results demonstrate that SCDM significantly reduces the interference between detection and ReID tasks, thereby enhancing tracking accuracy and efficiency.