Abstract: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.

3D Multi-Object Tracking with Adaptive Cubature Kalman Filter for Autonomous Driving

Object tracking with 3D LIDAR via multi-task sparse learning

Probabilistic 3D Multi-Object Cooperative Tracking for Autonomous Driving via Differentiable Multi-Sensor Kalman Filter

Enhanced Kalman Filter with Dummy Nodes and Prediction Confidence for Bipartite Graph Matching in 3D Multi-Object Tracking

HybridTrack: A Hybrid Approach for Robust Multi-Object Tracking

Probabilistic 3D Multi-Modal, Multi-Object Tracking for Autonomous Driving

A New Adaptive Noise Covariance Matrices Estimation and Filtering Method: Application to Multi-Object Tracking

A Tracking-By-Detection Based 3D Multiple Object Tracking for Autonomous Driving

Moving Camera-Based Object Tracking Using Adaptive Ground Plane Estimation And Constrained Multiple Kernels

A Novel Adaptive Noise Covariance Matrix Estimation and Filtering Method: Application to Multiobject Tracking

UKF‐MOT: an Unscented Kalman Filter‐based 3D Multi‐object Tracker

MCTrack: A Unified 3D Multi-Object Tracking Framework for Autonomous Driving

Online Multi-Target Tracking for Maneuvering Vehicles in Dynamic Road Context

Monocular Quasi-Dense 3D Object Tracking

3D Multi-Object Tracking in Point Clouds Based on Prediction Confidence-Guided Data Association

BiTrack: Bidirectional Offline 3D Multi-Object Tracking Using Camera-LiDAR Data

Three-dimensional Tracking of a Large Number of High Dynamic Objects from Multiple Views using Current Statistical Model

An IMM-Enabled Adaptive 3D Multi-Object Tracker for Autonomous Driving

Deformable Model-Based Vehicle Tracking and Recognition Using 3-D Constrained Multiple-Kernels and Kalman Filter

Multi-vehicle Detection and Tracking Based on Kalman Filter and Data Association

An online multi-object tracking approach by adaptive labeling and kalman filter