Generalizable Stable Points Segmentation for 3D LiDAR Scan-to-Map Long-Term Localization

Mobile robots increasingly operate in real-world environments that are subject to change over time. Accurate and robust localization is, however, crucial for the effective operation of autonomous mobile systems. In this paper, we tackle the challenge of developing a generalizable learned filter for long-term localization based on scan-to-map matching, using only 3D LiDAR data. Our primary objective is to enhance the reliability of mobile robot localization in dynamic environments. To obtain a strong generalization capability of the learned filter, we exploit the discrepancy between scan and map data. Our approach involves applying sparse 4D convolutions on a joint sparse voxel grid that encompasses both, scan voxels and their corresponding map voxels. This allows us to segment scan points into stable and unstable points based on a predicted long-term stability confidence score for each scan point. Our experimental results demonstrate that utilizing the stable points for localization improves the performance of scan-matching algorithms, especially in environments where changes in appearance are frequent. By exploiting the discrepancy between scan and map voxels, we enhance the segmentation of stable points. As a result, our approach generalizes to new, unseen environments