A contrastive Siamese association learning approach for 4D radar multiobject tracking

Abstract

Robust multi-object tracking is essential for autonomous driving systems, where accurate perception of dynamic objects directly supports trajectory prediction, obstacle avoidance, and path planning. Compared with cameras and LiDAR, 4D millimeter-wave radar provides all-weather sensing capability and direct Doppler velocity measurements, making it a promising sensor for reliable tracking in challenging environments. However, 4D radar point clouds are typically sparse, noisy, and temporally unstable, which makes identity-consistent association difficult, especially when nearby targets interact or radar reflections fluctuate across frames. To address these challenges, this paper proposes a contrastive Siamese association framework for 4D radar multi-object tracking. Object-level point clouds from historical tracks and current observations are mapped by a full-weight-sharing Siamese encoder into compact dense embeddings, where discriminative point-cloud correlations are learned for track-observation matching. The learned embedding matching probability is further fused with a constant-velocity position likelihood to construct a unified association matrix for online data association. In addition, a hybrid training objective with contrastive warmup and progressive binary cross-entropy mixing is combined with range truncation and static-point filtering to improve training stability, suppress clutter, and enhance score calibration. Experiments on V2X 4D radar data under an oracle-detections protocol demonstrate that the proposed embedding-motion fusion strategy reduces identity switches and improves AMOTA and IDF1 compared with a JPDA-style motion-only baseline. These results indicate that learned radar point-cloud correlation can effectively complement interpretable motion cues, leading to more robust association for 4D radar multi-object tracking.