Uncertainty-Driven Spatiotemporal Matching for Large-Scale LiDAR Relocalization

Yue Yao

Abstract: Estimating a reliable 6-DoF global pose from LiDAR point clouds remains challenging in dynamic and structure-poor environments. Current regression techniques typically treat consecutive scans as independent inputs, ignoring the inherent sequential continuity of vehicle motion. This work introduces a methodology that explicitly models inter-scan relationships to improve temporal-aware localization stability. The approach predicts dense global coordinate estimation for individual points while simultaneously estimating their predictive uncertainty. To bridge information across time, we utilize an attention-based mechanism TempLoc that establishes soft correspondences between adjacent frames. These temporal priors are then fused with the single-frame prior coordinate generation through an uncertainty-guided coordinate fusion scheme, allowing the network to discount unreliable regions dynamically. Evaluated on the Oxford RobotCar and NCLT datasets, the method achieves significant improvements in translation and rotation accuracy compared to single-scan baselines, confirming that integrating multi-frame temporal context is crucial for robust global positioning.