A Robust Normal Estimation Algorithm Based on Statistical Distance

Normal vector of point cloud has being widely used in the field of laser sensor mapping, stereoscopic vision and surface reconstruction. Because of the present of noise, classical method based on locally plane fitting could not get an accurate result and greatly decrease precision of the follow-up work. This paper proposes a robust method for normal estimation in dealing with point cloud contained noise point. We first obtain the best set, which have the maximum consistency, using the difference of statistical distance between inliers and outliers, then introduce median and the Median Absolute Deviation to remove noise point from the best set, finally get the locally best-fit-plane. Experiment results show that our method could efficiently couple with samples containing 50% noises and get accurate normal vectors. This new method is of great value in surface reconstruction, point cloud characterization, segmentation, matching or other reverse engineering task.