Unsupervised Segmentation Of Urban 3D Point Cloud Based On Lidar-image

Abstract

As one of the most vital fields in 3D scene understanding, segmentation of lidar point cloud has verity of applications such as unmanned vehicle. In order to deal with the problem of limited local computing resources, this paper proposes a two-step unsupervised learning-free method to segment sparse 3D point cloud with low computational demands. The 3D point cloud data is projected into a lidar-specific 2D coordinate system called lidar-image. For unmanned vehicles, road is a special kind of object, so the algorithm divides road firstly by using improved lidar-histogram. The algorithm further examines the geometric connection between adjacent points in 3D space, thus dividing non-road points into independent subsets. Experiments on KITTI dataset validate the proposed algorithm outperforms similar methods.