SGE-GLoc: Semantic Gaussian Ellipsoid Scene Graphs for Efficient LiDAR Global Localization

Abstract

Global localization, encompassing robust place recognition and precise transformation estimation, is crucial for mobile robot navigation when the global navigation satellite system (GNSS) is unavailable. While LiDAR-based approaches are favored for their accuracy in 3D perception and resilience to illumination changes, existing methods often suffer from complex scene representations or computationally intensive processes, hindering scalability in large-scale applications. This paper introduces SGE-GLoc, a novel LiDAR-based global localization framework, which utilizes a compact scene graph representation where individual objects are modeled as Semantic Gaussian Ellipsoids (SGEs), encapsulating both geometric and semantic attributes. For efficient place recognition, we encode SGE features into histograms and employ a Graph Neural Network (GNN) to learn local and global topological information, generating a discriminative two-stage descriptor for scene similarity scoring. For accurate 6-DoF transformation estimation, we first establish initial correspondences based on SGE matching and subsequently refine the transformation by exploiting geometric constraints between SGEs. Extensive evaluations on four public datasets, including KITTI, KITTI-360, Apollo, and HeLiPR, demonstrate that SGE-GLoc achieves high place recognition accuracy and low localization error, while maintaining a highly compact scene graph representation with an average storage of only 1.28 KB per scene. Furthermore, the integration of SGE-GLoc into a SLAM system showcases its practical efficacy in real-world scenarios.