A Dynamic 3D Multi-Object Tracking Method Based on Spatiotemporal Features
Hui Li,Haoran Yang,Xiaoxue Ai,Zhong Chen,Yanli Wu
DOI: https://doi.org/10.1109/tiv.2024.3508743
IF: 8.2
2024-01-01
IEEE Transactions on Intelligent Vehicles
Abstract:3D multi-object tracking is one of the important research directions in computer vision and holds significant research value in the field of autonomous driving. Relying solely on single image information or point cloud information is insufficient to overcome tracking challenges in complex scenarios. Currently, multimodal fusion 3D tracking methods still face numerous issues in fusion performance, data association, and trajectory management. Therefore, this paper proposes a dynamic 3D multi-object tracking method based on spatiotemporal features. First, a multi-scale spatial feature embedding fusion network is designed to enhance the weight of critical information within different modal features, thereby improving the prominence of target features. Second, a temporal aggregation embedding module is proposed to address the characteristics of point cloud features and fusion features, enhancing feature alignment when target features are integrated into temporal features, resulting in more robust temporal features. Finally, a multi-stage hybrid affinity dynamic association module and an adaptive dynamic trajectory management module are combined to reduce the impact of similar targets on tracking, which improves the model's ability to perceive target positions in dense scenes, and enhances the robustness of target association matching. Experimental results on the KITTI dataset have demonstrated that the proposed method achieves better tracking performance compared to other state-of-the-art methods.