Robust Outlier Detection and Saliency Features Estimation in Point Cloud Data

This paper investigates outlier detection and reliable local saliency features (e.g. normal) estimation in point cloud data. We propose two highly robust outlier detection algorithms that are able to identify outliers and are efficient for reliable local saliency features estimation in noisy point cloud data. One is based on a univariate robust z-score and the other on a multivariate Mahalanobis type robust distance. They combine the ideas of orthogonal distance and local surface points consistency to get Maximum Consistency with Minimum Distance (MCMD). Experimental results are presented to show the algorithms´ performance and are compared with other existing methods for synthetic and real datasets through segmentation for planar and non-planar surfaces of complex objects. The algorithms give more accurate and robust results, are fast and have the potential for local surface reconstruction, fitting, registration and covariance statistics based point cloud processing.