Semantic-Rich 3D CAD Models for Built Environments from Point Clouds: An End-to-End Procedure

The last few years has been subject to an unprecedented growth in the application of 3D data for representing built environment and performing engineering analyses such as site planning, and condition assessment. Despite the growing demand, generating semantic rich 3D CAD models, particularly when building and structural systems are exposed remains a labor-intensive and time-consuming process. To address these limitations, this paper presents an end-to-end procedure to produce semantic rich 3D CAD models from point cloud data at a user defined level of abstraction. The procedure starts by segmenting a point cloud while considering local context using a multi-scale region growing algorithm. A Markov-Random-Field optimization labels segments based on their semantic categories. This step reduces the over-segmentation produced during the segmentation stage by compositing similarly labeled segments into super segments. The interconnectivity among these super-segments are reasoned and b-splines and solid geometrical representations are fit to produce 3D NURBS surfaces and cylindrical elements, respectively. Experimental results on real-world point clouds show an average fit error of 6.33 E-01 mm making the method the first to include beams and columns in an automated Scan2BIM process.