Boundary constrained voxel segmentation for 3D point clouds using local geometric differences

Abstract

In 3D point cloud processing, the spatial continuity of points is convenient for segmenting point clouds obtained by 3D laser scanners, RGB-D cameras and LiDAR (light detection and ranging) systems in general. In real life, the surface features of both objects and structures give meaningful information enabling them to be identified and distinguished. Segmenting the points by using their local plane directions (normals), which are estimated by point neighborhoods, is a method that has been widely used in the literature. The angle difference between two nearby local normals allows for measurement of the continuity between the two planes. In real life, the surfaces of objects and structures are not simply planes. Surfaces can also be found in other forms, such as cylinders, smooth transitions and spheres. The proposed voxel-based method developed in this paper solves this problem by inspecting only the local curvatures with a new merging criteria and using a non-sequential region growing approach. The general prominent feature of the proposed method is that it mutually one-to-one pairs all of the adjoining boundary voxels between two adjacent segments to examine the curvatures of all of the pairwise connections. The proposed method uses only one parameter, except for the parameter of unit point group (voxel size), and it does not use a mid-level over-segmentation process, such as supervoxelization. The method checks the local surface curvatures using unit normals, which are close to the boundary between two growing adjacent segments. Another contribution of this paper is that some effective solutions are introduced for the noise units that do not have surface features. The method has been applied to one indoor and four outdoor datasets, and the visual and quantitative segmentation results have been presented. As quantitative measurements, the accuracy (based on the number of true segmented points over all points) and F1 score (based on the means of precision and recall values of the reference segments) are used. The results from testing over five datasets show that, according to both measurement techniques, the proposed method is the fastest and achieves the best mean scores among the methods tested. (C) 2020 Elsevier Ltd. All rights reserved.