StreamMOTP: Streaming and Unified Framework for Joint 3D Multi-Object Tracking and Trajectory Prediction

• DOI: 10.48550/arxiv.2406.19844
• Publication Date: 2024-06-28
• Authors: Jiaheng Zhuang; Guoan Wang; Siyu Zhang; Xiyang Wang; Hangning Zhou; Ziyao Xu; Chengqi Zhang; Zhiheng Li
• Keywords: Tracking (education)

3D multi-object tracking and trajectory prediction are two crucial modules in autonomous driving systems. Generally, the tasks handled separately traditional paradigms a few methods have started to explore modeling these joint manner recently. However, approaches suffer from limitations of single-frame training inconsistent coordinate representations between tasks. In this paper, we propose streaming unified framework for Multi-Object Tracking Prediction (StreamMOTP) address above challenges. Firstly, construct model exploit memory bank preserve leverage long-term latent features tracked objects more effectively. Secondly, relative spatio-temporal positional encoding strategy is introduced bridge gap maintain pose-invariance prediction. Thirdly, further improve quality consistency predicted trajectories with dual-stream predictor. We conduct extensive experiments on popular nuSences dataset experimental results demonstrate effectiveness superiority StreamMOTP, which outperforms previous significantly both Furthermore, also prove that proposed has great potential advantages actual applications driving.