A Neural Predict-Update Framework for Multi-Object Tracking Using Partial Observations

Multi-object tracking (MOT) aims to detect targets continuously while preserving identity consistency and estimating their dynamic states. Conventional approaches typically employ the Kalman Filter (KF) and its variants as the core state estimator, yet their performance heavily depends on precise motion modeling and manual noise tuning, which often deteriorates in complex or sensor-impaired environments. To address these limitations, we propose a 3D MOT framework built upon a Neural Prediction-Update (NPU) module. The NPU replaces the KF's prediction step with a lightweight neural network and adopts an observation-centric update strategy, thereby eliminating the need for explicit motion models and improving robustness against uncertain dynamics. Furthermore, to extend our framework to scenarios with only 2D detections, we propose a Projection-guided Physical-constraint-based Matching (PPM) method that searches for 3D states whose projections best match the detected 2D boxes. By enforcing geometric and physical consistency, PPM generates reliable 3D observations, improving spatial estimation under partial sensing conditions. Extensive experiments on the KITTI dataset demonstrate that our framework achieves superior accuracy and robustness compared with existing 3D MOT methods.