Dual-Graph Based Particle Filter Matching for Global Localization in Large-Scale Indoor Environments

Abstract

Global localization in 3D maps relies on distinctive object landmarks and geometric cues from structural elements. However, the presence of repetitive structures (e.g., long corridors, identical rooms) poses significant challenges for precise map matching due to scene ambiguity. Probabilistic models have been employed to mitigate the uncertainty. Yet, they still tend to fall into local optima when observations are insufficient. In this letter, we propose a dual-graph based particle filter matching method to address the problems. Unlike previous work that simply discards structural features to reduce perception ambiguity, our method organizes structural features into a separate structural graph, along with the conventional object graph, and incorporates them into a probabilistic pose estimation framework. We evaluate the method on simulated, self-collected, and publicly available real-world indoor datasets. Overall, Dual-Graph reduces the average accumulated online convergence time by 53.3% compared with the fastest convergent baseline, improves the average stable success rate from 43.9% to 81.6%, and achieves 97.57% precision, versus 94.35% for the state-of-the-art method.