6 days ago
9
Electronics, Vol. 13, Pages 4950: Enhanced Kalman Filter with Dummy Nodes and Prediction Confidence for Bipartite Graph Matching in 3D Multi-Object Tracking
Electronics doi: 10.3390/electronics13244950
Authors: Shaoyu Sun Chunyang Wang Bo Xiao Xuelian Liu Chunhao Shi Rongliang Sun Ruijie Han
Kalman filter (KF)-based methods for 3D multi-object tracking (MOT) in autonomous driving often face challenges when detections are missed due to occlusions, sensor noise, or objects moving out of view. This leads to data association failures and cumulative errors in the update stage, as traditional Kalman filters rely on linear state estimates that can drift significantly without measurement updates. To address this issue, we propose an enhanced Kalman filter with dummy nodes and prediction confidence (KDPBTracker) to improve tracking continuity and robustness in these challenging scenarios. First, we designed dummy nodes to act as pseudo-observations generated from past and nearby frame detections in cases of missed detection, allowing for stable associations within the data association matrix when real detections were temporarily unavailable. To address the uncertainty in these dummy nodes, we then proposed a prediction confidence score to reflect their reliability in data association. Additionally, we modified a constant acceleration motion model combined with position-based heading estimation to better control high-dimensional numerical fluctuations in the covariance matrix, enhancing the robustness of the filtering process, especially in highly dynamic scenarios. We further designed bipartite graph data association to refine Kalman filter updates by integrating geometric and motion information weighted by the prediction confidence of the dummy nodes. Finally, we designed a confidence-based retention track management module to dynamically manage track continuity and deletion based on temporal and reliability thresholds, improving tracking accuracy in complex environments. Our method achieves state-of-the-art performance on the nuScenes validation set, improving AMOTA by 1.8% over the baseline CenterPoint. Evaluation on the nuScenes dataset demonstrates that KDPBTracker significantly improves tracking accuracy, reduces ID switches, and enhances overall tracking continuity under challenging conditions.
