Automation pipeline and its implementation from point cloud to building information modeling

Building Information Modelling (BIM) is nowadays the most commonly used practice in the design and documentation of buildings. However, the creation of BIM is a very manual process and requires a considerable amount of time even for a skilled modeler. The primary focus of this research is designing an automation pipeline from captured scan data to BIM. The proposed pipeline can assist or fully replace modelers in creating BIM, and help companies digitize existing buildings with higher efficiency and lower cost. The pipeline is further broken down into four different processes to facilitate automation. The input is limited to point cloud, the most common data format for building imaging technology. The input point cloud is first cropped, voxelized, and aligned with the coordinates. The room-oriented point cloud clustering algorithm strategically adopts both 2D image processing techniques and the 2D cell complex to separate the input into rooms. The results are further verified with a customized user interface. The structure-oriented point cloud segmentation and classification are achieved in this step. Two novel neural networks are designed to facilitate the point cloud semantic and instance segmentation. The BEACon network aims to define a clearer separation between the instance boundary, while the BIMNet focuses on better efficiency and removing unnecessary calculations. With a classified point cloud, a room surface model is generated. A series of carefully designed mesh cleaning and simplification algorithms are developed to ensure that the output is rectified, simple, and contains as much detail as possible. Lastly, the structural BIM components are reconstructed from the surface model. Using prominent rules such as parallelism and rectangularity, the geometry is accurately captured, together with the topology relationship between each of the components. This dissertation presents a machine learning based pipeline for Scan-to-BIM conversion, which can handle multimodal and noisy data input of arbitrary scale. The resulting as-built BIM model includes all major structural components with topological information. The pipeline also includes a cloud deployment with a user-friendly interface and can reduce the time and effort required for BIM creation through fast processing iterations. One of the key innovative contributions of this work is the use of BIMNet, a neural network that can generate semantic and instance information for each point cloud data point. This network is an improvement upon previous hierarchical learning approaches for point clouds and demonstrates enhanced efficiency and robustness. The use of BIMNet in the BIM system has the potential to greatly impact the speed and accuracy of Scan-to-BIM conversion, and could potentially revolutionize the field of BIM. Additionally, The development of the complete Scan-to-BIM pipeline represents a significant advance in the field and has the potential to greatly streamline the process of BIM creation. The pipeline includes innovative techniques such as adaptive voxelization and Polyfit-based surface reconstruction and is implemented in a complete system that represents a significant engineering achievement. Limitations and potential future works are also discussed in the last part of this dissertation.Doctor of Philosoph