The roof plane is a fundamental element in urban building structures. The precise segmentation of building roofs is crucial for reconstructing 3D models based on laser point cloud data. Currently, roof plane segmentation is mainly achieved by using point cloud plane segmentation algorithms, which can be divided into two categories: point-based and voxel-based methods. Point-based methods are not limited by shape and density and have strong applicability and high robustness; however, their segmentation accuracy depends on the feature estimation accuracy. Although voxel-based algorithms can calculate features more accurately and efficiently, they are more sensitive to the voxel grid size. The roof plane types of urban buildings include but are not limited to arched, sloping, and folded roofs. Diverse shapes and uneven distributions are major challenges for precise roof plane segmentation, which is one of the main reasons for over- and under-segmentation. Irrelevant patches generated by oversegmentation and unclear planar boundaries caused by undersegmentation lead to unsatisfactory segmentation accuracy. Accordingly, we propose a novel airborne LiDAR point cloud roof plane segmentation method based on voxel and point hybrid growth. The advantages of our method are as follows: 1) a more accurate feature estimation, 2) the reduced generation of pseudoplanes at the boundaries, and 3) an effective allocation of plane intersection points. The experimental results demonstrate the validity of the proposed approach. We hope that this project can provide a new idea for 3D point cloud data processing and analysis as well as provide technical support for many tasks, such as urban surveying and planning and high-precision urban 3D model reconstruction.

Roof-plane segmentation is a crucial step in the 3D reconstruction of buildings. In this study, we propose a novel airborne LiDAR point cloud roof plane segmentation algorithm based on voxel and point hybrid growth. The proposed approach comprises three parts (Fig. 1). In the first part, voxels are separated into planar and non-planar voxels via an octree structure, and the feature information is then calculated for each planar voxel. In the second part, we design a plane coarse segmentation based on the voxel and point hybrid growth. Continuity, coplanarity, and distance features are the main constraints in the growth stage for obtaining the initial planar patches. Finally, in the third part, fine plane segmentation based on the cumulative distance voting mechanism is proposed to redistribute the unallocated points. The final segmentation results can be obtained by using a simple merging operation, according to the continuity and coplanarity constraints.

This study proposes a novel voxel and point hybrid growth algorithm for roof plane segmentation. The proposed method has three advantages. 1) Our approach uses voxels as the minimum feature computation unit. Hence, it can obtain more precise normal and curvature information, compared with point-based methods. As shown in Fig. 4, larger voxels are mostly found in smooth regions, whereas smaller voxels are found at the edges, corners, or around the rough regions. This implies that local surface planarity can be represented well by smaller voxels in the margins. 2) We design the residual value and point cloud number of voxels as criteria for the seed voxels, which not only solves the problem of the improper selection of seed voxels but also avoids the generation of pseudoplanes at the boundaries. 3) A novel voxel and point hybrid growth mechanism is proposed to allocate points at the plane junction, which can reduce the impact of voxel size changes. A reasonable allocation of boundary points can also yield better segmentation results. The experimental results (Figs. 15 and 16) indicate that the proposed method performs well for diverse roof-plane segmentations. In particular, for roofs with multiple planes, our method can also distinguish boundaries effectively and obtain high accuracy (Table 2), compared with other state-of-the-art approaches (Figs. 17 and 18).

In this study, we present a novel approach for roof-plane segmentation from airborne LiDAR point clouds. The proposed voxel and point hybrid growth algorithm can accurately calculate the feature information and reasonably assign boundary points. First, the voxels are separated via an octree structure, which can distinguish the voxels at the junctions of planes to accommodate local planarity. Second, we employ continuity, coplanarity, and distance constraints in a voxel and point hybrid region growth model to acquire the initial planar patches. Finally, a cumulative distance voting mechanism is used to reassign unsegmented points to improve the accuracy. The segmentation results are obtained after a simple merging operation. The experimental results prove that the aforementioned algorithms are valid and feasible. Compared with the traditional RANSAC, region growth, and EVBS methods, the proposed approach appears to have more advantages in complex roof plane segmentation. Finally, our study can be used for 3D model reconstruction, urban surveying, and planning.