ReLiTrack: Cross-Sensor 3D Tracking via Observation-Transition Factorization

Autonomous-driving fleets do not keep one sensing configuration throughout their lifetime. LiDAR placement, sensor composition, return policy, sampling density, coordinate processing, and annotation cadence change as platforms and data pipelines evolve. The resulting archives are expensive to recollect, yet a tracker trained on one configuration may bind object appearance, motion scale, and confidence to that configuration, making historical point-cloud sequences difficult to reuse on a new vehicle. We study this data-reuse problem for LiDAR-based 3D single-object tracking, where each localization error also changes the next search region and can therefore accumulate over time. We introduce ReLiTrack, a single-state tracker that separates a history-only, continuous-time transition prior from acquisition-sensitive point observations. The model constructs masked point tokens without duplicating sparse measurements, compares the current crop with a permanent initialization anchor and recent valid memories, and predicts one appearance innovation together with its uncertainty. Source observations are paired with range–density-degraded views to discourage configuration-specific evidence shortcuts, while covariance intersection combines the motion and appearance estimates without assuming independent errors. A frozen all-direction protocol uses multiple public driving datasets only as distinct sensing configurations for testing whether one source checkpoint can be reused across platforms.