Multi-Object Tracking Based on Long and Shortterm Trajectory Management

Multi-object tracking, capable of providing valuable information for decision-making systems, serves as a critical component of autonomous navigation and plays a highly significant role in the field of autonomous driving. In traditional 3D point cloud multi-object tracking methods, the tracking accuracy is generally low due to incorrect target associations or premature trajectory terminations. That is, if the predicted trajectory fails to be successfully associated and matched with newly detected targets within a short period, the predicted trajectory will be terminated. This leads to frequent switches of the target ID, and further results in incorrect tracking of the targets. This problem is particularly evident in complex dynamic environments where there are numerous targets, target overlaps, and mutual occlusions among targets. To tackle the above issues, with the intention of reducing the number of target ID switches, lowering the incorrect tracking rate, and enhancing the tracking accuracy, this paper proposes a 3D point cloud multi-object tracking method based on long and short - term trajectory management. Finally, the method of this paper is applied to the nuScenes dataset and the campus measurement dataset. The quantitative and qualitative experimental results show that the method of this paper has improved in the four indexes Of AMOTA AMOTPMOTA and IDS compared with the existing 3D point cloud multi-object tracking methods. which verifies the effectiveness of the method of this paper.