3D multi-object tracking with boosting data association and improved trajectory management mechanism

In the multi-object tracking (MOT) algorithm based on the tracking-by-detection paradigm, matching accuracy between detection and prediction, robustness to occlusion and consistency between trajectory and ground truth are three emphasized issues. In this study, we propose a tracking-by-detection-based online 3D MOT framework, which produces 3D bounding boxes with identity, position, and shape in real time from point clouds provided in each sequential frame. Specifically, first, to increase the correct matching rate, we propose a 3D efficient affinity (3D-EAFF) that combines position correlation and geometric features to optimize data association between predictions and actual detections. Second, to reduce the identity switches caused by occlusion, a trajectory deletion strategy is proposed by using multiple distance thresholds to judge occlusion as well as setting a decay coefficient to eliminate prediction error accumulation. Third, to enhance the smoothness of the trajectory, a trajectory update strategy based on the position relationship between detection and prediction is proposed. By reallocating weights between measurements and prior estimates for Kalman filter updates, we aim to achieve a more optimal dynamic balance. The experimental results on two representative 3D MOT benchmarks NuScenes and KITTI show that our method effectively improves the tracking accuracy, and its robustness to occlusion also reaches competitive performance.