LM-Mapping: Large-Scale and Multi-Session Point Cloud Consistent Mapping

In the field of autonomous driving and mobile robotics, constructing high-precision prior maps is a significant problem. For large outdoor scenes, maps often need to be segmented or collected repeatedly. Issues such as sensor degradation and measurement error can result in map inconsistencies or local precision loss. To solve this problem, we propose a new method for processing large-scale point cloud maps: LM-mapping. In this method, the point cloud map is divided into multiple different map blocks, each constructing a pyramid-like multiple constraints from bottom to top. The bottom constraint enhances the local accuracy of the map, while the top constraint addresses the joint registration of multi-session point cloud maps. Both sets of constraints are simultaneously integrated into the factor graph for optimization, enabling the system to address the challenges of local accuracy loss and the inconsistency in multi-session point cloud maps concurrently. To further reduce memory occupation and time consumption, we propose an adaptive feature voxel in the process of point cloud joint optimization. This method removes point clouds in non-feature areas in advance, thereby reducing unnecessary space division. Additionally, we refine the voxel division method. Experiments show that our method is more efficient and reasonable. we verify our method on data above 60 km, and the experimental results fully prove that our method can solve the problems of local accuracy loss and multi-session map inconsistency simultaneously.