Min-Cut Based Segmentation of Airborne Lidar Point Clouds

Abstract

ABSTRACT:Introducing an organization to the unstructured point cloud before extracting information from airborne lidar data is common in many applications. Aggregating the points with similar features into segments in 3-D which comply with the nature of actual objects is affected by the neighborhood, scale, features and noise among other aspects. In this study, we present a min-cut based method for segmenting the point cloud. We first assess the neighborhood of each point in 3-D by investigating the local geometric and statistical properties of the candidates. Neighborhood selection is essential since point features are calculated within their local neighborhood. Following neighborhood determination, we calculate point features and determine the clusters in the feature space. We adapt a graph representation from image processing which is especially used in pixel labeling problems and establish it for the unstructured 3-D point clouds. The edges of the graph that are connecting the points with each other and nodes representing feature clusters hold the smoothness costs in the spatial domain and data costs in the feature domain. Smoothness costs ensure spatial coherence, while data costs control the consistency with the representative feature clusters. This graph representation formalizes the segmentation task as an energy minimization problem. It allows the implementation of an approximate solution by min-cuts for a global minimum of this NP hard minimization problem in low order polynomial time. We test our method with airborne lidar point cloud acquired with maximum planned post spacing of 1.4 m and a vertical accuracy 10.5 cm as RMSE. We present the effects of neighborhood and feature determination in the segmentation results and assess the accuracy and efficiency of the implemented min-cut algorithm as well as its sensitivity to the parameters of the smoothness and data cost functions. We find that smoothness cost that only considers simple distance parameter does not strongly conform to the natural structure of the points. Including shape information within the energy function by assigning costs based on the local properties may help to achieve a better representation for segmentation.

Review of 3D Point Cloud Data Segmentation Methods

“…Compared with several segmentation methods, the results show that this method is suitable for segmenting foreground and background, suitable for specified target extraction, or implementing multi-objective extraction in a supervised classification manner. Ural et al [28] suppose that a conditional random field model exists, and use graph cut algorithm to disconnect part of the graph model to form independent regions according to the maximum posterior estimation criterion. Li et al [29] proposed a progressive form of a two-level optimal segmentation algorithm.…”

Two-Layer-Graph Clustering for Real-Time 3D LiDAR Point Cloud Segmentation

“…These methods have made a lot of contributions to the point cloud storage structure. The use of a graph structure [17,18] enables the vast data to represent the relationship between different points and make full use of the storage space. However, the segmentation standard designed by the above methods is relatively simple, and in some cases, over-segmentation occurs.…”

Min-Cut Based Semantic Building Labeling for Airborne Lidar Data

“…Research employing MRFs for point classification often employ point features derived from the eigenvalues of the covariance matrix of points within a neighborhood. Ural and Shan (2012) label points as surface or scatter via min-cut optimization using point features derived from the eigenvalues of the structure tensor. Similarly, Sun and Salvaggio (2013), classify trees by graph-cut optimization for which the data costs are calculated by thresholding the smallest eigenvalue.…”