Density-guided and topology-constrained road boundary extraction from urban scene point clouds

Road boundary extraction remains a challenging task due to the structural complexity and variability of real-world road environments. In this paper, we propose a density-guided and topology-constrained road boundary extraction method that jointly exploits local geometric cues and global structural relationships to achieve accurate and robust boundary detection. First, ground points are extracted by applying a region growing algorithm based on the geometric features of supervoxels. Then, a height difference operator based on point-wise z-values is designed to eliminate non-boundary points. To highlight boundary regions, a hyperbolic density response (HDR) is formulated to amplify relative density contrast through a nonlinear transformation. By incorporating global topological constraints, the initial curb points are segmented to compute curb-wise center points, from which a topology-consistent road centerline is optimized by enforcing directional continuity between adjacent segments. Finally, road boundary geometries are categorized into straight segments, curved segments, and outliers according to the angular relationship between boundary orientation and centerline direction. Adaptive optimization strategies are subsequently applied to different boundary types. Experimental results on two datasets collected using different scanning systems demonstrate that the proposed method consistently achieves robust and accurate road boundary extraction across diverse environments, yielding precision, recall, and F1-score of 0.97, 0.95, and 0.96 on the Paris-Lille-3D dataset, and 0.98, 0.97, and 0.97 on the WHU-TLS dataset, respectively.