A fast segmentation method of sparse point clouds

Abstract

In this paper, we present a fast segmentation algorithm based on the geometric characteristics of the objects and the attribute of medium. This algorithm is not only suitable for sparse point clouds, but also for dense point clouds. It is built up of three stages: First, the range image is established from the Velodyne VLP-16 laser scanner data, which changes the sparse characteristic of data in the original space and determines the close relationship between the data points. Then, according to the geometric relation of the adjacent data points and point clouds edges distribution analysis, a region growing method is used to complete the fast segmentation of point clouds data, obtaining a series of mutually disjoint subsets. Finally, based on the laser intensity, refined segmentation of the under-segmentation subset is addressed using the K-means clustering method. The point clouds of an indoor corridor scene are used to verify the superiority of our method and compared with three typical algorithms. Experimental results prove that our method can fastly and accurately segment objects in the scene, and is not sensitive to noise and satisfactory in anti-noise performance.