Objective With the rapid development of virtual reality technology, indoor navigation technology, and indoor positioning technology, the extraction and modeling of indoor 3D point cloud objects have become a research hotspot. Under normal circumstances, an indoor scene is quite complex, and the point cloud data obtained by scanning is usually cluttered. There are many objects and occlusions, and automatic modeling cannot be carried out. It is necessary to segment a complex indoor point cloud into simple geometric primitives to perform modeling. Because there are several plane structures in indoor scenes, such as walls and ground, plane segmentation for indoor scene point cloud is a crucial part of segmentation for indoor scene point clouds. Owing to the complexity and bulkiness of indoor scene point clouds, traditional random sample consensus (RANSAC) and 3D Hough transform methods are complex and inefficient in the process of plane segmentation for indoor scene point clouds. In this article, we propose a new method for plane segmentation of indoor scene point clouds. Compared with existing methods, this method has a great improvement in time efficiency and is more suitable for plane segmentation for indoor scene point clouds.
Methods In this article, a new plane segmentation method based on projection length point cloud layering and mean shift (MS) normal vector constraint is proposed. First, the method estimates the normal vector of the point cloud by the principal component analysis method, combines the coordinates of the point cloud to obtain the projection length, and then layers the point cloud according to the projection length by a certain step. Afterward, it takes the current maximum stratified point cloud for normal vector constraint based on the MS method to get the point cloud with the most concentrated normal vector. Next, it uses the remaining points to perform RANSAC and least squares plane fitting to obtain the plane parameters and then removes the point cloud contained in the current plane model by a certain thickness threshold. The above steps are repeated to obtain the parameters of all planes until the number of plane points extracted is less than a certain value. Finally, the model point clouds are extracted from the original point cloud based on the obtained plane parameters, and after the model optimization that includes planes merging, error point reclassification, and irrelevant point elimination, the final plane segmentation result is obtained.
Results and Discussions In this article, a new concept of projection length of point cloud is proposed that is used to segment the plane of point cloud in an indoor scene (Fig. 2). The indoor point cloud is layered on the basis of the projection length, and the resultant point cloud number histogram can initially reflect the number of planes and distance distribution in the scene (Fig. 4). The projection lengths of the planes calculated from the resulting plane parameter fall in the peak or adjacent interval in the resultant point cloud number histogram (Table 2). After the point cloud layering based on the projection length, most points in the maximum layer come from the same target plane, and there are only a small number of irrelevant points. After MS clustering, the remaining points are all from the target plane, which is convenient for plane fitting (Fig. 5). The proposed method can completely segment the plane structure of indoor scenes, including walls, ceiling, floor, and desktop. Meanwhile, other irrelevant structures, such as potted plants, chairs, and door frames, are removed in the segmentation process (Figs. 7 and 8). The distances between the obtained plane models are very close to the actual measured distances; the difference is in the millimeter level (Table 3). The deflection angles between the planes obtained in this study, and the planes obtained by single-point measurement are all within 0.2° (Table 4). Compared with the maximum likelihood sample consensus method and improved 3D Hough transform method, the proposed method is obviously better in terms of total time consumption (Table 5).
Conclusions In this article, we propose a new method of plane segmentation for indoor scene point cloud. Through the point cloud layering based on projection length and normal vector constraint based on MS, the proposed method can quickly obtain points from a single plane, thereby achieving plane fitting and segmentation rapidly and then gets the final result after model optimization. Experiments show that the proposed method can effectively segment the plane structure in the indoor scene point cloud, and the model optimization can avoid over-segmentation and remove irrelevant points. Simultaneously, the experiment proves that the segmentation result of the proposed method has higher accuracy and meets the requirements of later modeling. In addition, compared with two improved classical methods for point cloud segmentation, the proposed method is time efficient and is suitable for segmentation for a large number of point clouds.
