TripletLoc: One-Shot Global Localization Using Semantic Triplet in Urban Environments

Abstract

This study presents a system, TripletLoc, for fast and robust global registration of a single LiDAR scan to a large-scale reference map. In contrast to conventional methods using place recognition and point cloud registration, TripletLoc directly generates correspondences on lightweight semantics, which is close to how humans perceive the world. Specifically, TripletLoc first respectively extracts instances from the single query scan and the large-scale reference map to construct two semantic graphs. Then, a novel semantic triplet-based histogram descriptor is designed to achieve instance-level matching between the query scan and the reference map. Graph-theoretic outlier pruning is leveraged to obtain inlier correspondences from raw instance-to-instance correspondences for robust 6-DoF pose estimation. In addition, a novel Road Surface Normal (RSN) map is proposed to provide a prior rotation constraint to further enhance pose estimation. We evaluate TripletLoc extensively on a large-scale public dataset, HeliPR, which covers diverse and complex scenarios in urban environments. Experimental results demonstrate that TripletLoc could achieve fast and robust global localization under diverse and challenging environments, with high memory efficiency.