Research On Normal Estimation And Isosurface Extraction For Implicit Surface Reconstruction Of 3D Point Cloud

With the rapid development of science and technology,three-dimensional scanning technology and processing technology have made great breakthroughs.Especially in recent years,on account of the availability of low-cost 3D scanning equipment,3D point clouds of sampled real-world objects are widely disseminated.At the same time,the ultimate purpose of most data acquisition is still to reconstruct the surface similar to the real object.Among the many types of reconstruction methods,implicit surface reconstruction is widely used due to its effective processing of noise,missing data and non-uniform sampling.However,in practical applications,as the complexity of the processed objects increases,how to reconstruct the surface accurately and rapidly has become a challenge.Therefore,this thesis conducts in-depth research on the normal estimation,consistent orientation and rapid isosurface extraction in the process of implicit surface reconstruction:Considering the normal estimation of point cloud,a robust and accurate local plane voting method based on entropy is proposed.Our method presupposes that the normal of a point can be constituted from at least one in the normal set from the planes composed of triples in the neighborhood of this point,and its core idea is a local plane voting strategy.Each voting takes the entropy value of the current plane and the credibility of the plane into consideration.In addition,the average fitting residuals of the neighborhood and the plane density weights are designed to participate in local plane voting,which can further effectively deal with noise and non-uniformly sampled point clouds.Tests on a large number of synthetic and real scan point clouds show that our method can generate high-quality normals and is robust to noise and non-uniform sampling.Aiming at the consistent orientation of the normal of the point cloud,a minimum spanning tree construction method based on bilateral weights is proposed,and an adaptive flip criterion is utilized to further ensure that the propagation errors of the normal orientation of each point are minimized.The method focuses on consistent orientation issue of unorganized point clouds with complex surfaces,especially close-by and/or sharp surfaces,and stresses on enhancing the association of the same region and the stable propagation between different regions.It mainly includes: on the one hand,an optimized priority-driven normal propagation scheme is introduced,where lower possibility across different regions;On the other hand,an improved flip criterion can ensure reliable normal direction.Plentiful experiments show that the method can generate consistent orientation even under various difficult scenarios,which provides a prerequisite for high-quality implicit surface reconstruction.Proceed from addressing the time consumption caused by the efficiency issue of isosurface extraction in implicit surface reconstruction of point cloud,an automatic isosurface growth method based on voxel vertex state is proposed.The proposed approach avoids scanning the entire voxel space,and can automatically search for boundary voxels without additional conditions,thus speeding up the extraction process of isosurface.In addition,the method allows execution from arbitrary or specified voxel positions and its framework can be applied to other related isosurface extraction algorithms.Experiments show that this method can remarkably improve the extraction efficiency.