Robustness of LiDAR-Based Pose Estimation: Evaluating and Improving Odometry and Localization Under Common Point Cloud Corruptions

TL;DR

The paper addresses the challenge of robust LiDAR-based pose estimation under realistic point cloud corruptions, covering both odometry and localization. It introduces a framework to systematically evaluate five SOTA pose-estimation systems across 18 synthetic corruptions and investigates mitigation strategies, including denoising and corruption-based re-training for learning-based methods. Findings show that odometry systems are vulnerable to specific corruptions, while localization-based methods like LocNDF remain robust, and that corruption-aware training substantially boosts robustness and can improve performance on clean data. The work provides practical guidance for improving LiDAR pose estimation in adverse conditions and shares code and data to enable reproducibility and further research.

Abstract

Accurate and reliable pose estimation, i.e., determining the precise position and orientation of autonomous robots and vehicles, is critical for tasks like navigation and mapping. LiDAR is a widely used sensor for pose estimation, with odometry and localization being two primary tasks. LiDAR odometry estimates the relative motion between consecutive scans, while LiDAR localization aligns real-time scans with a pre-recorded map to obtain a global pose. Although they have different objectives and application scenarios, both rely on point cloud registration as the underlying technique and face shared challenges of data corruption caused by adverse conditions (e.g., rain). While state-of-the-art (SOTA) pose estimation systems achieved high accuracy on clean data, their robustness to corrupted data remains unclear. In this work, we propose a framework to systematically evaluate five SOTA LiDAR pose estimation systems across 18 synthetic real-world point cloud corruptions. Our experiments reveal that odometry systems degrade significantly under specific corruptions, with relative position errors increasing from 0.5% to more than 80%, while localization systems remain highly robust. We further demonstrate that denoising techniques can effectively mitigate the adverse effects of noise-induced corruptions, and re-training learning-based systems with corrupted data significantly enhances the robustness against various corruption types.

Figures (3)

• Figure 1: An example of rain and wet ground corruption. The field view and reflection intensities are significantly decayed compared to the clean data. The color bar indicates the intensities of the points.
• Figure 2: Performance of each system on corrupted KITTI LiDAR Odometry Sequence 09 and 10. Errors with significantly different magnitudes from the majority are aligned at their magnitude and the actual values are labeled in the figure. For example, an error of 12.79% is aligned to "$>$10.0". If multiple errors overlap at the same magnitude, we label the two highest severity levels.
• Figure 3: (a) Performance on the denoised data. (b) Performance of the retrained model.