Abstract: Automatic generation of building information model (BIM) from point cloud data has been a hot topic in building automation. The disadvantages of automatic 3D building reconstruction based on traditional algorithms include manual design features, complex identification process and limited application scenes. As 3D machine learning matures, there are new ways to deal with point clouds. The point cloud was processed by introducing the network ASIS in instance segmentation, that is, the architectural construction elements were automatically segmented and classified from the scanning of point cloud scene and the instance segmentation matrix was obtained. Then, the external contour parameters of building components were extracted from the obtained instance segmentation matrix based on the bounding box hypothesis, and the external contour parameters and the semantic classification results of segmentation were taken as component parameters of BIM. Finally these extracted component parameters were input into the self-made IFC generator to automatically generate the BIM model based on the Industry Foundation Class (IFC) standard. Experiments show that the method of noise-free point cloud can realize 3D reconstruction of indoor single room based on the hypothesis of Manhattan world.

Key words: deep learning, industry foundation class, building information modeling, automation; point cloud