DTNLS: 3D Point Cloud Segmentation Based on 2D Image and 3D Point Cloud Double Texture Feature

Panoramic segmentation of 3D point clouds is an essential and challenging technology for robots with 3D detection and measurement capabilities. In order to fuse the color information of 2D image pixels with the spatial position information of the 3D LiDAR point cloud, it is necessary to establish the corresponding relationship between the RGB of pixels and the XYZ position of the 3D LiDAR point cloud. We present Double Texture Neighbor LiDAR Segmentation (DTNLS) in this letter. Double texture refers to the color texture of the image and the point cloud texture of the 3D LiDAR. The DTNLS method first uses the color texture feature of the image to segment the pixels and then obtains the clustering center and segmentation boundary outline. 3D LiDAR point cloud texture segmentation takes the clustering center obtained above as the diffusion source, diffuses outwards along the ring LiDAR line, finds the LiDAR point cloud texture boundary features near the image segmentation boundary contour, and finally realizes 3D point cloud segmentation. We carried out quantitative analysis experiments, and the results showed that compared with the best results among the existing mainstream segmentation methods, the proposed DTNLS method improved the accuracy of pedestrian segmentation by 32.2%. The recall improved by 20.12% in pedestrian segmentation.
IoU
improved by 48.8% for pedestrians. We conducted empirical studies on public datasets and our datasets to demonstrate that DTNLS has broad applicability and better performance in 3D point cloud segmentation than the previous latest techniques, without the need for any 2D images and 3D point cloud training data.