Diagnostics based Principal Component Analysis for Robust Plane Fitting in Laser Data

Abstract

Plane fitting and obtaining characteristics (e.g., normal) from the estimated plane are fundamental tasks in many applications in which laser scanner 3D data is used. Unfortunately, laser data are not free from outliers. Principal Component Analysis (PCA) is a popular method for plane fitting, but it is known that PCA is very sensitive to outliers and gives misleading non-robust results. We present a robust plane fitting algorithm based on PCA coupled with an outlier detecting diagnostic statistical approach. In this method, the recently introduced robust scatter matrix is used to calculate robust statistical distance for finding outliers. After excluding outliers, PCA is performed on the outlier free data which is used for fitting planar surfaces and to estimate robust normal and other parameters. Demonstration of the new algorithm through several synthetic and vehicle based laser scanning data show that the proposed method is efficient, and gives robust estimates. Results outperform Least Squares (LS), PCA and are significantly better than the well-known RANSAC in terms of time, accuracy and robustness. This method has great potential for robust segmentation, surface reconstruction, and other point cloud processing tasks.