3D Multi-Object Tracking for Autonomous Driving Based on IMM-EKF and Re-Identification

For the 3D multi-object tracking (MOT) task in autonomous driving, frequent identity switches (IDS) may lead to traffic accidents. For the IDS issue of the tracked objects, we propose a 3D MOT algorithm based on extended Kalman filter of interacting multiple model (IMM-EKF) and re-identification (ReID). Our method improves tracking performance while reducing the IDS number of the tracked objects. We have improved the following three aspects. First, we remove low confidence and repeated 3D detection results to reduce false matches. Second, we use the IMM-EKF method to predict and update the tracking trajectories. IMM can combine the advantages of multiple motion models to perform state fusion estimation, and EKF can effectively handle the nonlinear motion issue of the tracked objects. Third, we adopt a two-stage association matching. The first stage association matching based on spatial position and the second stage association matching based on the ReID features can increase the matching accuracy. On the nuScenes validation set, compared with the baseline algorithm, our method improves AMOTA by 4.4%, MOTA by 3.8%, and reduces the IDS number by 213.