3D Object Segmentation Considering Density Variation and Scanning Order of Point Clouds

In this paper, we propose a novel DBSCAN method for 3D object segmentation that considers point cloud density based on lidar range and scanning order of laser beams. Point clouds obtained by a lidar are dense at short ranges and sparse at long ranges. Furthermore, point clouds obtained by the lidar are organized in the scanning order of each beam. However, conventional DBSCAN has issues that its parameters are fixed and cannot adapt to density variation of point clouds, and that the computational costs for neighbor- hood search are high, resulting in significant processing time. Therefore, in the proposed method, parameters are adjusted according to lidar ranges to accommodate density differences. Furthermore, executing neighborhood search within only a specific beam scanning area reduces computational costs and shortens processing time. We conducted experiments in multiple environments to verify the effectiveness of the proposed method. By automatically determining parameters based on the point cloud density, the proposed method demonstrated that both nearby and distant objects can be correctly segmented. Addi- tionally, our method demonstrated that processing time can be reduced by executing neighborhood search within only a specific scanning area of the beams. As described above, the proposed method achieves adaptive and high-speed object segmentation by considering both the differences in point cloud density due to lidar ranges and the beam scanning order.