LearnTrack: Learning-Based Uncertainty Modeling for Kalman Filter Optimization in LiDAR-Based Multiobject Tracking

Abstract

Accurate uncertainty characterization is essential for recursive state estimation in dynamic measurement systems. The Kalman filter (KF) has served as the predominant solution for processing measurement signals in such tasks. However, the KF assumes independent and identically distributed noise, which is frequently violated in practical applications. Existing approaches struggle to capture real measurement characteristics, rendering state estimation suboptimal. This article addresses this limitation through learning-based uncertainty modeling directly from measurement data. Two signal processing methodologies are systematically compared: the proposed optimized KF (OKF) that implicitly absorbs the coupled disturbance while maintaining linear structure, and a neural KF (NKF) that explicitly compensates unmodeled dynamics through recurrent architectures, serving as a representative data-driven baseline. A novel Track-Level Observation-Ground Truth (GT) Alignment strategy enables supervised learning from existing benchmarks. LiDAR-based 3-D multiobject tracking (MOT) serves as the case study, where measurement uncertainties are highly state-dependent. Both methodologies are instantiated through LearnTrack—a unified framework. LearnTrack-L integrates the OKF with detection-confidence modulation (DCM); LearnTrack-N incorporates the NKF for complex motion modeling. Experiments on nuScenes and KITTI datasets demonstrate that the OKF outperforms both conventional KF and NKF in trajectory precision with reduced estimation drift. In 3-D MOT, LearnTrack-L improves tracking precision while significantly reducing identity switches (IDSs) by 28. These results confirm that supervised noise parameter learning offers a practical approach to enhance state estimation performance in real measurement systems. The source code is available at https://github.com/Still-Wang/LearnTrack