Semantic-Geometric Triple Constraint-Based Graph Matching for Robust Multi-Robot Global Localization in Large-Scale Environments

Vision-based multi-robot global localization in large-scale environments is a challenging task due to real-time environmental changes, which complicate data association between viewpoints. Recently, researchers have proposed methods to enhance viewpoint invariance by organizing object semantic information using graph structures. However, previous works still face limitations in accuracy and robustness in real-world scenarios, primarily due to the limited descriptive power of feature information (semantic or geometric) and susceptibility to noise. In this paper, we propose a novel semantic-geometric triple constraint-based graph matching multi-robot global localization method, called SGT-MGL. We first extract the semantic and geometric features of objects and describe the distribution of neighboring objects using topological structures. To improve object discriminability, we construct a triangular descriptor for each object based on semantic information and relative distances. Considering the complementarity between semantic and geometric information, we introduce a 3D histogram descriptor that encodes semantic, spatial angular, and relative distance information, enhancing the invariance of the triangular descriptor. To further mitigate noise, we propose a candidate point selection strategy guided by global geometric structures and employ a combined local and global graph matching approach for 6-DOF pose estimation. We extensively evaluate SGT-MGL on three public datasets, demonstrating superior accuracy and robustness. The implementation of SGT-MGL will be available.