Research

Map matching is critical for location-based applications. However, there remain several gaps to satisfy stringent engineering requirements. First, the GPS trajectory data sparsity problems exist due to the expensive data annotation costs, which limits the development of data-driven map matching methods. Second, the ubiquitous noise brings the challenge to the stability performance of map matching algorithms.

This research develops an end-to-end trajectory representation learning method for offline map-matching with diffusion-based trajectory generation. Based on a diffusion model, the TrajDiffuse algorithm is proposed for trajectory generation.And a trajectory representation learning model MapformerX is developed for map matching. Four main components  in MapformerX are utilized to extract the internal relationship of trajectory sequences as well as the external relationship between trajectory sequences and segment roads. Empirical studies using both the real-world GPS trajectory dataset Planet-GPX-GPS and synthetic datasets show that the proposed method outperforms start-of-the-art baselines for  offline map-matching and maintains robustness under noisy datasets.