Point Cloud Denoising Research Based On L₀ Normal Optimization And Feature Rectification

3D point cloud is a basic representation of 3D objects,which is realized by sampling continuous 3D objects into discrete points.Usually,various sensors can be used to directly obtain point cloud data from the real world.With the continuous progress of 3D scanning technology,the increasing popularity of various 3D sensing devices,and the simple,flexible and powerful representation ability of 3D point cloud itself,3D point cloud has become an increasingly popular 3D representation form,appearing in various fields.However,due to the limitations of 3D point cloud acquisition equipment and the interference of various environmental factors during the acquisition process,the collected point cloud data is inevitably contaminated by noise.The specific performance is that the coordinate information of the point space in the point cloud is disturbed,deviates from the surface of the body,and cannot accurately restore the geometric information of the three-dimensional object.At the same time,many subsequent downstream tasks,such as 3D reconstruction,segmentation and classification,require clean and accurate point cloud data.Therefore,point cloud denoising is a very important preprocessing process.For the problem of 3D point cloud denoising,many methods have been proposed.For example,from the earliest methods based on filtering and optimization,and the gradually developed methods based on data-driven deep learning.The denoising method is a trade-off between noise removal and feature details preservation.While removing noise points,preserving sharp features is also one of the most challenging problems.The main work and innovation of this thesis is to propose an optimization method based on L0 norm of normal vector,which optimizes the normal first-order difference and second-order difference,and can adapt to different types of geometric surfaces.At the same time,aiming at the problem of preserving sharp features,a local dihedral angle frame is proposed for feature extraction and normal correction.The following are the key steps of this article:1.Due to the point cloud model polluted by noise,the point normal vector calculated according to the point position distribution will also be disordered.In order to correct the point normal vector of each point and adapt to different types of geometric surfaces in the point cloud model,the method of combining the first-order difference optimization of the point normal vector with the second-order difference optimization of the point normal vector is introduced.The second order difference optimization of point normal vector can make the transition of point normal vector in point cloud model smoother.At the same time,the first-order difference optimization of the point normal vector can keep the geometric features of the flat region as much as possible.2.After the point normal vector optimization and pre-denoising process,the original sharp feature areas in the point cloud will be smoothed to different degrees,so that the sharp features will no longer be maintained.To solve this problem,this thesis proposes a sharp feature rectification method based on local dihedral angle frame.Through the properties of the local dihedral angle frame,the sharp feature points and their nearby points are screened.Then,according to the geometric features of sharp feature points and their adjacent points,the sharp feature rectification is realized.Finally,a large number of detailed experimental results show that the method in this thesis can effectively remove the noise in the 3D point cloud model,and can also restore the geometric structure of the sharp feature areas in the point cloud.