Global Localization based on Road-centric 3D Point Cloud Descriptor in Urban Environments

Abstract

Global localization is essential for unmanned vehicle navigation in a prebuilt map especially in weak or erroneous GNSS signals areas. Unlike images containing rich features, point clouds are sparse and almost contain pure geometric information which makes global localization using a single scan still a challenging problem. Based on Scan Context framework, we propose a road-centric 3D point cloud descriptor for global localization in urban environment. Rather than generate the descriptor in the raw sensor coordinate, we incorporate road measurements and topological information to transform the point cloud into a road-centric coordinate and generate the descriptor which makes it theoretically rotation-invariant and translation-invariant. In addition, we can obtain more precise initial pose to align point cloud with the prebuilt map through the transformation between raw sensor coordinate and road-centric coordinate. Evaluation on KITTI dataset shows that our proposed method has better performance in global localization and the success rate is 26% higher than Scan Context.