Coupling Point Cloud Completion and Surface Connectivity Relation Inference for 3D Modeling of Indoor Building Environments

Abstract

Due to occlusions and limited measurement ranges, three-dimensional (3D) sensors are often not able to obtain complete point clouds. Completing missing data and obtaining spatial relations of different building components in such incomplete point clouds are important for several applications, for example, 3D modeling for all objects in indoor building environments. This paper presents a framework that recovers missing points and estimates connectivity relations between planar and nonplanar surfaces to obtain complete and high-quality 3D models. Given multiple depth frames and their sensor poses, a truncated signed distance function (TSDF) octree is constructed to fuse the depth frames and estimate the visibility labels of octree voxels. A normal-based region growing method is utilized to detect planar and nonplanar surfaces from the octree point cloud. Based on the surfaces and the visibility labels, missing points are completed by estimating the connectivity relations between pairs of the surfaces and by filling individual planar surfaces. Experimental results demonstrate that the proposed method can correctly identify at least 78% of the connectivity relations between the detected surfaces, and 87% of added points are correct and help to generate high-quality 3D models compared to the ground truth model.