Automated Semantic Segmentation of Indoor Point Clouds from Close-Range Images with Three-Dimensional Deep Learning

Abstract

The creation of building information models requires acquiring real building conditions. The generation of a three-dimensional (3D) model from 3D point clouds involves classification, outline extraction, and boundary regularization for semantic segmentation. The number of 3D point clouds generated using close-range images is smaller and tends to be unevenly distributed, which is not conducive to automated modeling processing. In this paper, we propose an efficient solution for the semantic segmentation of indoor point clouds from close-range images. A 3D deep learning framework that achieves better results is further proposed. A dynamic graph convolutional neural network (DGCNN) 3D deep learning method is used in this study. This method was selected to learn point cloud semantic features. Moreover, more efficient operations can be designed to build a module for extracting point cloud features such that the problem of inadequate beam and column classification can be resolved. First, DGCNN is applied to learn and classify the indoor point cloud into five categories: columns, beams, walls, floors, and ceilings. Then, the proposed semantic segmentation and modeling method is utilized to obtain the geometric parameters of each object to be integrated into building information modeling software. The experimental results show that the overall accuracy rates of the three experimental sections of Area_1 in the Stanford 3D semantic dataset test results are 86.9%, 97.4%, and 92.5%. The segmentation accuracy of corridor 2F in a civil engineering building is 94.2%. In comparing the length with the actual on-site measurement, the root mean square error is found to be ±0.03 m. The proposed method is demonstrated to be capable of automatic semantic segmentation from 3D point clouds with indoor close-range images.