Abstract: Point cloud deep learning technology was used to automatically identify the dimensions of building components. Virtual point clouds were used to solve the problem of cumbersome work with 3D point cloud datasets. First, a method for creating a self-built point cloud dataset was proposed, which used building information modeling (BIM) tech-nology to parametrically model fabricated components. The data was then batch processed and converted to generate a noise-free, high-quality point cloud with annotations. Next, based on the PointNet network, an end-to-end component size parameter identification net-work, termed PointNet CE, was built. Finally, the virtual point cloud dataset was used for model training, and the effectiveness of the method was verified via engineering examples. The experimental results showed that the virtual point cloud dataset generated based on BIM technology could effectively expand the real world data. The improved component size parameter recognition network could accurately identify the component size, with a recognition accuracy at millimeter level on the training sample. This level of accuracy was suitable for real world use, with the recognition accuracy of the components reaching centimeter level, which satisfied the construction requirements of the fabricated structure.

Key words: building information modeling (BIM), fabricated component, point cloud, deep learning, virtual dataset