Abstract
Map matching is a fundamental technology that links positioning data to high-precision road networks, yet it continues to face significant challenges in noise robustness, cross-regional transferability, and interpretability. Existing methods often struggle with local-global feature fusion, adaptation to dynamic road network changes, and reliance on opaque black-box models. To address these limitations, this paper introduces SceneGTMM, a transferable GNN-Transformer dual-graph interaction framework for map matching, built upon a conformal mapping-based scene-relative strategy. The proposed framework comprises three key components: (1) a conformal mapping-based scene-relative strategy that constructs trajectory-centric local coordinate systems, reducing dependence on the training road network and enabling robust cross-regional transfer and seamless adaptation to dynamic road network updates; (2) a GNN-Transformer dual-graph interaction architecture, where a GNN-modeled road graph captures local topological constraints and a Transformer-modeled trajectory graph captures global temporal dependencies, with cross-graph attention mechanisms facilitating noise suppression and semantic alignment; and (3) a CRF-enhanced structured prediction module that integrates the Transformer's global context with CRF's topological transition constraints, improving path connectivity and overall robustness. Experimental results demonstrate that SceneGTMM achieves over 80% accuracy on multi-source trajectories with positioning errors ranging from 16 to 50 meters, representing a 5.3% improvement over traditional HMM-based approaches. In cross-city transfer scenarios, SceneGTMM outperforms existing state-of-the-art methods including MTrajRec, GraphMM, and TMM, while also enhancing interpretability through attention visualization and relative coordinate analysis. This work establishes a new paradigm for high-precision, transferable map matching, with direct applications in real-time traffic perception and autonomous driving path planning.
via ArXiv CV