LiDAR-Based 3D Multi-Object Tracking with VoxelNeXt

Tracking-by-Detection (TBD) paradigm-based 3D multi-object tracking (MOT) with LiDAR sensors have gained widespread attention in computer vision and autonomous driving due to their efficiency and portability. TBD paradigm decomposes 3D MOT into two sequential stages: detection and data association. Thus, the accuracy of TBD-based algorithms is jointly influenced by the performance of the detector and association strategies. To validate the effectiveness of data association in TBD, we measured four association strategies based on the VoxelNeXt detector. Our experiments demonstrate that the VoxelNeXt-based 3D MOT framework works best with a hierarchical data association strategy. VoxelNeXt with hierarchical data association strategy delivers outstanding tracking performance on the nuScenes benchmark compared to state-of-the-art methods, achieving a 2.9% improvement in AMOTA over baseline approaches.