Research On Normal Vectors Estimation Method For Unstructured Point Clouds

Recently,physical digital technology has been applied in reverse engineering,medical visualization and many other domains.Because the collected surface data is dense and unstructured,it is called “point clouds”.As fundamental property of 3D shape,normal vector is a basic condition of various algorithms,and it has an extensive application in point clouds processing.For instance,many surface reconstruction algorithms require the normal vectors as input;during feature extraction,the normal vector is an important tool to describe the geometric shape of objects;also,in point cloud segmentation,the variation of normal vectors can be helpful to identify the boundaries.In this paper,we mainly improve the existing algorithms for some perspectives,including algorithm robustness,computational efficiency,sharp features processing and surface variation measurement based on normal vectors.The majority contributions of our work are summarized below:(1)We propose a surface Variation measurement criterion(NV)based on normal vector: firstly,we estimate normal vectors of point clouds by traditional methods,and then use quadric surface to fit local normal vectors distribution and construct target equations.Also,we solve the minimum values for the equations and get analytical solutions of surface variation.Experiments can prove that compared with surface variation criterion(SV),the physical meaning of NV criterion is clearer,and the distinction of surface variation described by NV is better.(2)We propose a mixed normal vector estimation strategy for point clouds: ?1 we propose the multi-scale PCA algorithm.The algorithm adopts scale fusion method,and it gets final normal vectors by weighted fusion under different search radius.So it can improve the robustness for normal vector estimation under noise.?2 The point clouds are divided into two aspects like flat and feature regions by normal variation(NV).For flat region,we utilize Multi scale PCA,and for feature region,Hough CNN method is used to estimate normal vectors.So our mixed strategy effectively reduces the running consumption and improves the accuracy of normal vector estimation.(3)We propose the normal vector estimation algorithm based on depth features classification and neighborhood optimization for point clouds with sharp features.?1 We propose a normal vector algorithm that treats sharp features specially through discretizing the normal vector space and transforming the problem to feature classification,so it can reflect the normal mutation at sharp features.?2 A normal vector refining method is present,which uses the difference between the initial normal vectors to distinguish neighborhood points of different local surface patches,and obtains the exact normal vector from the refined neighborhood points.Our method can eliminate the discretization errors of initial normal vectors.?3 Eventually,we optimize the original Hough CNN network architecture and propose a lightweight feature mapping network-Sharp Net to get results of point clouds with features rapidly.