Abstract

ABSTRACT The objective of segmentation on point clouds is to spatially group points with similar properties into homogeneous regions. Segmentation is a fundamental issue in processing point clouds data acquired by LiDAR and the quality of segmentation largely determines the success of information retrieval. Unlike the image or TIN model, the point clouds do not explicitly represent topology information. As a result, most existing segmentation methods for image and TIN have encountered two difficulties. First, converting data from irregular 3-D point clouds to other models usually leads to information loss; this is particularly a serious drawback for range image based algorithms. Second, the high computation cost of converting a large volume of point data is a considerable problem for any large scale LiDAR application. In this paper, we investigate the strategy to develop LiDAR segmentation methods directly based on point clouds data model. We first discuss several potential local similarity measures based on discrete computation geometry and machine learning. A prototype algorithm supported by fast nearest neighborhood search and based on advanced similarity measures is proposed and implemented to segment point clouds directly. Our experiments show that the proposed method is efficient and robust comparing with algorithms based on image and TIN. The paper will review popular segmentation methods in related disciplines and present the segmentation results of diverse buildings with different levels of difficulty.