LiDAR Pose Estimation via Uncertainty-Aware Temporal Correspondence Modeling
Lists
Only you see this
Abstract
Global pose estimation from LiDAR scans is a fundamental challenge in outdoor autonomous navigation, where the goal is to determine the 6-DoF localization of a sensor within a prior map. Existing regression-based localization methods predominantly rely on single-scan inference, which fails to leverage inter-frame spatio-temporal relationships and degrades performance in ambiguous or dynamically changing environments. To address this limitation, we present a localization framework that explicitly exploits temporal continuity across consecutive LiDAR observations to achieve robust outdoor relocalization. Our framework consists of three key components: a global prior estimation module that generates global coordinate predictions together with per-point uncertainty estimates, a coordinate-guided generation module that employs transformer-based attention to model point-level associations between successive scans, and an uncertainty-aware coordinate fusion module that integrates multi-frame coordinate predictions in an end-to-end trainable pipeline. By jointly reasoning over spatial uncertainty and temporal context, TempLoc produces globally consistent 6-DoF pose estimates that are resilient to transient occlusions and scene ambiguities. Comprehensive evaluations on the NCLT and Oxford RobotCar datasets demonstrate that our method surpasses current state-of-the-art relocalization approaches by a substantial margin, validating the benefit of explicit temporal correspondence modeling for LiDAR-based global localization.
BibTeX
@misc{zhao2026lidar,
author = {Xibin Zhao},
title = {{LiDAR Pose Estimation via Uncertainty-Aware Temporal Correspondence Modeling}},
year = {2026},
url = {https://openreview.net/forum?id=oiIQ5dT6JE},
howpublished = {OpenReview},
}Loading related content...