Multi-Dimensional Semantic Map Design and High-Precision Localization for Train Group Control

Abstract

Group-based train control is expected to break the capacity bottleneck of conventional block-based signaling by operating trains in closely spaced formations, which requires continuous, precise train positioning and semantic maps that encode control-relevant elements such as the track centerline, slope, curvature, and speed limits. However, existing railway LiDAR SLAM systems produce unstructured point clouds that cannot be utilized directly by control functions. This paper presents a multi-dimensional semantic mapping and high-precision localization framework for train groups. The map is organized into geometric, semantic, topological, and control layers, in which control elements are stored in parametric form and shared across trains as element-level increments over vehicle-to-vehicle links with confidence-weighted fusion. Track geometry is estimated by fitting a B-spline centerline to semantically segmented point clouds with automatic degradation to geometry-only fitting when segmentation confidence is insufficient. The train state is estimated in a sliding-window factor graph fusing BDS-RTK, LiDAR, IMU, and stereo vision. Experiments on a synthetic line with full ground truth and onboard tests on the train show that the framework maintains decimeter level positioning in open sections, bounds the drift in GNSS-denied tunnels well below loosely coupled BDS/INS, and keeps track geometry estimation reliable under severe point cloud degradation. The results indicate that unstructured point clouds can be turned into control-ready map elements onboard, and that the geometry of these elements remains trustworthy even when the observations are severely degraded. Since the framework requires no trackside infrastructure beyond the BDS reference, it provides a practical basis for deploying group train control on existing lines.