Research On Feature-preserving Point Cloud Filtering

With the increasing popularity of consumer-level 3D scanners,computer graphics has advanced by leaps and bounds.Due to its property of containing rich geometric structure information,3D point clouds are becoming more and more prevalent.People use scanners to acquire 3D point cloud data to preserve 3D structural information of real-world objects.However,in addition to the accuracy of the scanner equipment,the surface of the scanned object is also uneven.Combined with various factors that affect the scanning quality,the scanned point clouds inevitably contain noise.Therefore,it becomes crucial to explore efficient and easy-to-use point cloud filtering algorithms.To address the aforementioned issues,this paper proposes a two-step method that is based on local manner.The method utilizes normal bilateral filtering with position updating and incorporates the repulsive force of the reference point distribution.This addresses the problem that local methods tend to aggregate points,which results in poor filtering quality.Secondly,a non-local and normal-independent point cloud filtering method is proposed in this paper.This method is able to obtain a filtered point cloud with a feature-preserving effect based on the point position only and solves the problems that the local method is prone to produce gaps and the method relying on the normal information has poor performance in maintaining the detailed features of the model.The main research of this paper is as follows.1.Local-based normal-guided point cloud filtering method.In order to address the problem that local-based point cloud filtering methods tend to aggregate points and cause poor quality of filtered point clouds,this paper first investigates point cloud filtering in a local manner and guided by normal information.To solve this problem,this study introduces a repulsive term for the reference point distribution in the geometric feature-aware point update technique,so that both point distribution and feature preservation are both considered in the filtering process.The key idea is to combine the repulsive term with the data term in an energy minimization process.The repulsive term is responsible for the point distribution,while the data term is used to project the noisy points onto the underlying surface while preserving the geometric features and obtaining the filtered point cloud.Experiments show that the method has achieved promising feature-preserving effect and produces relatively uniform point distribution.2.Nonlocal-based feature-preserving point cloud filtering method.To address the problems that local methods tend to generate gaps and unreliable normal information tends to distort the filtering results in the case of large noise,this paper then proposes a nonlocal point cloud filtering method that can maintain the geometric features of 3D models without normal information.The existing position-based point cloud filtering methods are difficult to preserve geometric features.This study rethinks point cloud filtering in a nonlocal and position-based manner.Then proposes a novel position-based point cloud filtering method for feature preservation.Unlike the normal-based technique,this method does not require normal information to preserve detailed features in the point cloud.The core idea is to first design a similarity metric to search for non-local similar patches of the local patch being queried.Then the non-local similar patches are mapped into a canonical space and the non-local information is aggregated.The aggregated result(i.e.,coordinates)will be in turn mapped to the original space as the final filtered point cloud.Extensive experiments validate our method,and show that it generally outperforms position based methods(deep learning and non-learning),and generates better or comparable outcomes to normal based techniques(deep learning and non-learning).3.Point cloud GUI visualization.In order to better display the point cloud processing results,this paper investigates how to combine point cloud filtering with GUI visualization.By using Open3D-GUI for point cloud visualization under the Python platform,integrating the existing point cloud processing algorithm with the point cloud filtering algorithm of this study in a module,and then developing a point cloud processing and visualization software platform.