Triplet-Graph: Global Metric Localization Based on Semantic Triplet Graph for Autonomous Vehicles

Global metric localization is one of the fundamental capabilities for autonomous vehicles. Most existing methods rely on global navigation satellite systems (GNSS). Some methods relieve the need of GNSS by using 3-D LiDARs. They first achieve place recognition with a pre-built geo-referenced point-cloud database for coarse global localization, and then achieve 3-DoF/6-DoF pose estimation for fine-grained metric localization. However, these methods require accessing point-cloud features and raw point clouds, making them inefficient and hard to be deployed in large-scale environments. To provide a solution to this issue, we propose a global metric localization method with triplet-based histogram descriptors. Specifically, we first convert the input LiDAR point clouds into a semantic graph and describe the vertices in the graph with the proposed descriptor for vertex matching and pose estimation. These vertex descriptors are then selected and aggregated into a global descriptor to decide whether two places correspond to the same place according to a similarity score. Experimental results on the KITTI dataset demonstrate that our method generally outperforms the sate-of-the-art methods.