Scan2BIM-NET: Deep learning method for segmentation of point clouds for scan-to-BIM

The architecture, engineering, and construction (AEC) industry perform thousands of scans each year. The majority of these point clouds are used for generating three-dimensional (3D) models - a process formally known as scan to building information modeling (Scan-to-BIM) - that represent the current conditions of a construction scene. Although point cloud data provide the scene's geometric information, its use presents several challenges that make the process of generating a 3D model from point cloud data time-consuming, labor-intensive, and error-prone. In order to address the mentioned challenges, this paper presents a new end-to-end deep learning method, named Scan2BIM-NET, for semantically segmenting the structural, architectural, and mechanical components present in point cloud data. It classifies beam, ceiling, column, floor, pipe, and wall elements using three main networks: two convolutional neural network (CNN) and one recurrent neural network (RNN). The method was trained and tested using 83 rooms from point cloud data representing real-world industrial and commercial buildings. The process returned an average accuracy of 86.13%, and the beam, ceiling, column, floor, pipe, and wall categories obtained an accuracy of 82.47%, 92.60%, 59.31%, 98.71%, 82.79%, and 84.46%, respectively. The experimental results showed that deep learning improves the accuracy of semantic segmentation of architectural, structural, and mechanical components. This new method has the potential of being a tool during the Scan-to-BIM process, especially for semantically segmenting underceiling areas where mechanical components are close to structural elements.