Unsupervised recognition of volumetric structural components from building point clouds

Forming semantically-rich 3D models of buildings from point clouds conveys vital as-built information that support applications such as construction quality assessment and control, construction progress tracking and infrastructure asset management. Conventional scan-to-BIM techniques rely on registration of scanned data with as-designed BIM objects. These methods are unable to handle cases where site-specific building element models are not readily available. Thus, this study proposes an unsupervised recognition framework to automatically identify generic building elements such as columns and beams from unstructured point clouds. The point cloud is first partitioned into multiple sections corresponding to each building floor. Next, column elements are identified using point cloud clustering and classification based on the bounding box dimensions. Beam elements are identified using plane projection and line fitting. The identified building elements are represented by volumetric bounding boxes in the resulting 3D model. The proposed method is validated using two separate datasets of laser-scanned building point clouds.