A long-term localization and mapping system for autonomous inspection robots in large-scale environments using 3D LiDAR sensors.

Inspection mobile robots equipped with 3D LiDAR sensors are now widely used in substations and other critical circumstances. However, the application of traditional LiDAR sensors is restricted in large-scale environments. Prolonged operation poses the risk of sensor degradation, while the presence of dynamic objects disrupts the stability of the constructed map, consequently impacting the accuracy of robot localization. To address these challenges, we propose a 3D LiDAR-based long-term localization and map maintenance system, enabling autonomous deployment and operation of inspection robots. The whole system is composed of three key subsystems: a hierarchical SLAM system, a global localization system, and a map maintenance system. The SLAM subsystem includes Local Map Representation, LiDAR Odometry, Global Map Formulation and Optimization, and Dense Map Generation. Specifically, we construct an efficient map representation that voxelizes only the occupied space and computes local geometry within each voxel. The design of LiDAR Odometry ensures high consistency with this map representation mechanism. Then, to address drift errors, we formulate the global map as a graph of local submaps that undergo global optimization. Furthermore, we utilize marching cubes to generate a mesh model of the map. Our system outperforms the state-of-the-art LiDAR odometry method, LOAM, reducing average absolute position error by 30 % and 38 % on two public datasets. The comparative evaluation highlights the system's superior accuracy and robustness, and demonstrates its high SLAM ranking in real-world scenarios. For global localization, we propose a novel ScanContext-ICP method, which integrates our improved ScanContext method, termed ScanContext++, for place recognition and global pose initialization. The Iterative Closest Point (ICP) algorithm is then employed for precise point cloud alignment and pose refinement, enabling the recovery of the robot's position on the offline map when localization is lost. Finally, the map maintenance system tracks environmental changes, distinguishing stable features from dynamic ones. The system assigns higher weight to stable voxels, thereby improving localization accuracy. Furthermore, our time distribution mechanism refines map updates by filtering unstable points through temporal and segment-level analysis, which further enhances map maintenance. We conduct extensive experiments on public datasets to validate our system. The experimental results demonstrate that our system is effective and can be deployed on inspection mobile robots.