An incremental lane-level network construction method based on high-precision trajectories

Lane-level network offers spatial references for autonomous driving, while trajectories remain crucial for their maintenance. However, prevailing methods exhibit limited ability in modeling network dynamics due to the intricate structure of traffic systems and the diverse driving behaviors. In this study, we introduce a three-stage framework for incrementally constructing a lane-level network by fusing trajectory data with existing map information. First, we design a binary classification model with specialized sequential and spatial features to detect and prune lane-changing trajectories, thus minimizing redundant topology construction. Subsequently, a lane-meanshift clustering algorithm is introduced that addresses spatial heterogeneity by applying a density-dependent penalty, integrating azimuth and distance in bandwidth selection to ensure robust candidate lane extraction across scenarios. Finally, lane-level partial map matching accurately detects new lane nodes. Post-processing then performs a redundancy examination, which refines the incremental construction and leads to iterative completion of the network. Comprehensive experiments demonstrate that our method outperforms state-of-the-art approaches on metrics for lane extraction, intersection topology extraction, and incremental update tasks. It constructs and updates accurate, complete lane-level networks in complex scenarios and exhibits strong robustness to variations in data volume and algorithm parameters.