Road Network Extraction of 3D Point Cloud Scene Based on L1-medial Extraction and Flexible Constraints

The 3D road network is crucial for urban traffic optimization, autonomous driving navigation, and disaster emergency response planning. However, obtaining it is highly challenging due to the complex and diverse environment. LiDAR point clouds, commonly used as a representation of 3D data, often suffer from disorganization and high noise levels, posing significant obstacles to road network extraction. This study proposes a method that effectively reduces the complexity of road network extraction by utilizing the L1 skeleton extraction method. Additionally, a symmetry-based approach is employed to construct a centroid distance field for the ground point cloud, facilitating erosion and constraining the central axis region of the road. The study detects the tail points of the initial road network skeleton and completes the road network based on this detection. Finally, flexible projection is used to determine the optimal central axis positions. Extensive experiments are conducted to evaluate the extracted 3D road network, demonstrating its effectiveness in extracting road network information from complex 3D roads and highlighting its significant advantages over 2D road networks.