Reliability-Aware LiDAR Relocalization across Partially Observed Routes

Hang Yu

Abstract: Scene Coordinate Regression has become an effective formulation for LiDAR-based global relocalization, yet its reliability depends heavily on whether the query trajectory is represented in the training data. In long-range autonomous-driving, this assumption is rarely satisfied: city-scale vehicles often enter route segments that were never observed during training, causing direct absolute pose regression to become unstable. We propose a reliability-aware relocalization framework for operation under partial scene coverage. The method uses an SCR backbone to produce global pose hypotheses and compact scene descriptors, then models temporal pose-feature consistency to estimate whether the current prediction should be trusted. A lightweight arbitration module selects between the predicted absolute pose and a propagated relative-motion estimate initialized from the most recent reliable state. This design maintains a globally referenced trajectory when the vehicle passes through uncovered regions, while retaining the accuracy of direct SCR in known areas. Experiments on UrbanUnseen and partial-coverage splits of Oxford RobotCar and NCLT show improved robustness over map-free relocalization baselines under partial coverage, particularly at known-to-unknown scene transitions.