A novel 3D point cloud segmentation algorithm based on multi-resolution supervoxel and MGS

Abstract

ABSTRACT As a basic and important research, point cloud segmentation plays an important role in many fields. However, the traditional point cloud segmentation algorithm still suffers from two problems. The first problem is that voxel-based segmentation algorithms cannot guarantee accuracy in regions of inconsistent density. The other problem is the inefficiency of point-based clustering algorithms. Hence, in order to solve the first two problems, a new supervoxel-based segmentation algorithm is proposed. To address the first problem, a multi-resolution supervoxel algorithm is proposed to obtain the basic unit for clustering, which includes a new low-density region detection algorithm and a resegmentation process. However, over-detection during the construction of supervoxels leads to the presence of small fragments around large supervoxels. Therefore, for the second problem, a novel BPSO (belief propagation supervoxel optimization) algorithm is proposed to optimize the supervoxel. Moreover, an improved multi-resolution supervoxel- and graph-based segmentation (MGS) algorithm is presented for supervoxel clustering and a segmentation optimization algorithm is adopted to allocate the unallocated points. Experiments are conducted on different datasets, and segmentation results are evaluated quantitatively. Compared with traditional methods and advanced methods, the results show that this method can segment urban point clouds accurately and effectively.