Global Registration of Point Cloud Maps with Low-overlap Regions

Merging multiple maps is necessary when the task of mapping a large area is divided and assigned to different robots. Unlike scanned or reconstructed models, point cloud maps built by SLAM are usually sparser and contain much noise, making current global registration methods failed. In this paper, we focus on the global registration of point cloud maps which have low-overlap regions and no prior relative transformation between each other, so that maps can still be merged in GPS-denied environment. Our method adopts a pipeline of feature matching, where key points are first selected evenly within voxel grids and then used for getting feature descriptors. After matching for correspondences, a RANSAC based registration gives the rough transformation between maps, which can further be refined by ICP. To improve the accuracy of correspondences, we specifically introduce a keypoint clustering and classification method. Except for 3D point cloud maps as inputs, no other information is needed. Experiments on the campus show that our approach successfully deals with large area with similar structures.