Region Proposal Mechanism for Building Element Recognition for Advanced Scan-to-BIM Process

Laser-scanned point clouds contain rich geometric information and are one of the primary modes of data acquired from as-built infrastructure. The creation of a building information model (BIM) from scanned data has several important applications such as construction progress tracking, asset management, structural health monitoring, quality assessment, and safety planning. A conventional scan-to-BIM pipeline typically involves a high volume of manual work to extract and label building elements which is time-consuming and error-prone. This research proposes a voxel grid search method to automatically generate region proposals for objects of interest from a laser-scanned point cloud. The advantages of the proposed method are that it does not require prior segmentation of objects and that it is agnostic with respect to the shape of objects that can be detected. A point cloud template is first created from a computer-aided design (CAD) model by randomly sampling points along the surfaces of the model. Next, a voxel grid data structure is used to represent the occupancy of both the template and the scene point cloud. A search procedure is used to determine candidate regions with high region proposal scores, and thresholding and mean shift clustering are used to refine the region proposal results. The proposed method was evaluated with a laser-scanned point cloud of a building, where a ground truth BIM model is available. Region proposals were generated for five different building elements (beam, column, roof, stairs, and window frame) and compared to the ground truth model using the metrics of precision, recall, and centroid error. Test results showed that the method achieved 0.04–0.50 for precision, 0.50–1.00 for recall, and 0.99–1.82 m for RMSE.