Robust methods for feature extraction from mobile laser scanning 3D point clouds

Abstract

Three dimensional point cloud data obtained from mobile laser scanning systems commonly contain outliers. In the presence of outliers most of the currently used methods such as principal component analysis for point cloud processing and feature extraction produce inaccurate and unreliable results. This paper investigates the problems of outliers, and explores advantages of recently introduced statistically robust methods for automatic robust feature extraction. The robust algorithms outperform classical methods and show distinct advantages over well-known robust methods such as RANSAC in terms of accuracy and robustness. This paper shows the importance and advantages of several recently introduced robust statistics based algorithms for (i) planar surface fitting, (ii) surface normal estimation, (iii) edge detection, and (iv) segmentation. Experimental results for real mobile laser scanning point cloud data consisting of planar and non-planar complex objects surfaces show the proposed robust methods are more accurate and robust. The robust algorithms have potential for surface reconstruction, 3D modelling, registration, and quality control for point cloud data.