Robust Ground Constrained SLAM for Mobile Robot With Sparse-Channel LiDAR

Although LiDAR SLAM has been widely studied over the past decades, it suffers from degradation situation, particularly for sparse-channel LiDAR. This study proposes a hybrid LiDAR SLAM method combining scan to LiDAR-centric sliding submap ICP and scan-to-map feature points matching to adapt to unstructured environments and degraded corridors. First, a ground constraint module based on the principal normal of the planar ground is proposed to avoid vertical drift. To address the lack of geometrical features in corridors and environments with large-scale featureless walls, intensity features combined with geometrical features are adopted for scan-to-map registration. Finally, a joint optimization module integrates the feature correlation, the IMU pre-integration, and the ground constraint into a factor graph to perform low-drift SLAM. Extensive experiments are conducted on the GroundRobotDataset, which include challenging degradation scenarios such as long corridor, obstacle crossing and unstructured outdoor environment with large-size walls in weak texture. Our method achieves better accuracy and robustness than state-of-the-art methods, particularly in inhibiting vertical drift.