Filling the Gap Between Using AI and Updating an Underground Network VECTOR Database: Skeleton Extraction from Classified Point Clouds

Antoine Carreaud, Fabio Mariani, Adrien Gressin

Abstract: This scientific study addresses the challenge of converting labeled 3D models into vector representations enriched with topology and attributes, which is a crucial step for updating vector databases such as underground networks. Relying on Convolutional Neural Networks for detection, the proposed skeletonization method employs a processing pipeline involving object separation, barycenter computation for point-like objects, and cylinder fitting for linear objects. The method ensures an accurate representation of the 3D model in vector format (points, lines, and polylines) with topology and attributes. Experimental validation, conducted on 200 excavation sites, attests to an accuracy of 85 % within a 10 cm threshold. Despite the limitations of partially recognized neighbors, the method was considered very satisfactory for updating a real underground network database in Geneva (Switzerland).