Point Clouds Outlier Removal Method Based on Improved Mahalanobis and Completion

Abstract

Point clouds have been regarded as a representative format for 3D visualization of real-world objects or scenes. However, point clouds acquired from depth cameras or laser scanning devices commonly contain outliers. Outlier removal performance will directly affect the downstream applications. Existing methods mainly perform outlier removal directly on the raw data, which are designed without finding a proper balance between outlier removal precision and detail feature retention. In this letter, we propose a novel outlier removal method called CIMD that can obtain outlier-free data by conducting the improved Mahalanobis distance and point clouds completion. A layered statistical outlier removal approach is introduced as preprocessing strategy to obtain feature points and incomplete point clouds. We filter feature points to fill the incomplete point clouds according to the improved Mahalanobis distances. The theoretical analysis proves that the improved Mahalanobis distance can magnify the difference between outliers and ground truth compared with the original way. Using publicly available dataset PointCleanNet and real-world scanned data, the experimental results show that the proposed method has superior performance compared with state-of-the-art methods.