Research On Point Cloud Normal Estimation Algorithm Based On Multi-scale Neighborhood Translation

In recent years,with the continuous popularization and development of 3D scanners,the gain of point cloud data has become very convenient.As an important feature of point clouds,the normal vector of point clouds plays a very major role in the rebuild,rendering,feature detection and extraction of point clouds.However,during the acquisition process of point clouds,objective factors such as noise and occlusion may affect the quality of the original point cloud data,resulting in a relatively large estimated normal error.Therefore,accurately estimating the normal vector of a point cloud is a relatively difficult task.In this thesis,a new point cloud normal estimation algorithm is proposed based on neighborhood translation as the core idea and Principal Component Analysis(PCA).PCA algorithm is mainly to fit all points in the neighborhood into a plane,and define the normal vector of the current point as the normal vector of the fitted plane.However,this algorithm constructs a candidate neighborhood centered on the current point.When the point cloud is located at special locations such as corner points and boundary points,its neighborhood is composed of multiple smooth surfaces,and the estimated normal error is large.The algorithm in this thesis includes more optional conditions in the construction of candidate neighborhood sets,providing a basis for selecting better neighborhoods.The evaluation criteria proposed in neighborhood filtering cover more aspects,making the selected optimal neighborhood more accurate.Therefore,the resulting normal direction is more accurate,which can effectively estimate the normal vectors of points located at boundaries,corners,and other locations,and has better robustness in handling non-uniform sampling problems.The steps of this algorithm are as follows: First,the initial normal vector of a point is estimated using the PCA algorithm,and then the degree of distance from all points to sharp features is characterized by the eigenvalues of the covariance matrix.The points are divided into smooth points and feature points;The second,to design two multiscale candidate neighborhood set generation methods based on all the nearest neighbors of the current point;Finally,two different optimal neighborhood evaluation methods are proposed to combine the "flatness" of the neighborhood with the "distance" from the current point.This article conducts different combination tests on the proposed algorithm.The results show that all combination tests are significantly superior to previous algorithms,and the time consumption is relatively low.However,from the perspective of the two schemes for constructing candidate neighborhood sets,the first scheme for constructing candidate neighborhood sets is better in terms of effectiveness,because the method filters more candidate neighborhoods;The second scheme is slightly less effective than the first scheme,but it consumes less time.