Clustermap building and relocalization in urban environments for unmanned vehicles

Abstract

Map building and map-based relocalization techniques are important for unmanned vehicles operating in urban environments. The existing approaches require expensive high-density laser range finders and suffer from relocalization problems in long-term applications. This study proposes a novel map format called the ClusterMap, on the basis of which an approach to achieving relocalization is developed. The ClusterMap is generated by segmenting the perceived point clouds into different point clusters and filtering out clusters belonging to dynamic objects. A location descriptor associated with each cluster is designed for differentiation. The relocalization in the global map is achieved by matching cluster descriptors between local and global maps. The solution does not require high-density point clouds and high-precision segmentation algorithms. In addition, it prevents the effects of environmental changes on illumination intensity, object appearance, and observation direction. A consistent ClusterMap without any scale problem is built by utilizing a 3D visual–LIDAR simultaneous localization and mapping solution by fusing LIDAR and visual information. Experiments on the KITTI dataset and our mobile vehicle illustrates the effectiveness of the proposed approach.