Point Cloud Segmentation Using Model-Fitting, Artificial Intelligence and Local Curvature Techniques

Semantic segmentation and point cloud classification within the context of Cultural and Architectural Heritage have become key topics of investigation in recent years, particularly due to advancements in Artificial Intelligence. While 3D data, acquired through methods such as laser scanning and photogrammetry, enable the generation of highly detailed representations of sculptures, buildings, and archaeological sites, they also present significant challenges in accurately distinguishing various architectural and structural components. This study proposes the development of an advanced methodological workflow, implemented in Python, which integrates well-established algorithms for model fitting (RANSAC), unsupervised learning as clustering (DBSCAN), and the analysis of geometric curvature within point clouds. The approach is fully automated and requires no manual pre-training, enabling the segmentation of elements such as pavement, benches, chairs, columns, vaults, and others through a clearly defined sequence of operations and precise parameter settings. The results obtained from this framework, applied to the study of the interior layout of an apulian church, are transferable to other case studies, adaptable to the specific needs of the user, and provide a solid foundation for future developments in computational representation, with potential applications in CAD or HBIM environments.