Global Localization for Single 3D Point Cloud using Voting Mechanism

Global localization on a given map is a vital problem for robot navigation tasks. Segments-based methods most rely on the dense point clouds, and do not work well when points are sparse. It needs robots to walk a certain distance to accumulate point clouds for segments, which is not safe for robots in an unknown environment. To solve this problem, we propose a novel global localization method which only needs the first single LiDAR scan at the initial stage when the robot starts. The first single LiDAR scan is treated as a query point cloud, the extracted descriptors of this query point cloud is compared with the prior Map’s descriptors in the database which are stored in a KD tree, and the most similar frame is selected for registration. In particular, we create a voting mechanism, a two-phase search strategy for place recognition, which reduces the query time. We evaluate our method on KITTI and MVSEC datasets, and our localization accuracy is increased by 52.8% compared with SegMap validated the effectiveness of our method.