Research And Platform Design Of 3D Reconstruction Technology Based On Point Cloud Data Processing

With the development of computer graphics,pattern recognition and other technologies,3D reconstruction technology has been widely used in mapping,unmanned driving,medical technology,robotics and other fields.3D reconstruction technology based on point cloud data processing uses laser scanner,structured light,stereo camera and other equipment to collect discrete point cloud data,and then analyzes,processes and optimizes these data,so as to achieve 3D reconstruction.In this respect,there are still many deficiencies in the research at home and abroad.For example,the calculation of point cloud filtering is large,the stability is poor,and the original features of 3D model are difficult to maintain.The registration algorithm needs to overcome many complex strategies such as noise,outlier,density change,partial overlap change,etc.And the accuracy of the model after the reconstruction of the surface is not enough,and the pseudo-surface is easy to appear.To solve these problems,a point cloud filtering method based on hybrid strategy and a point cloud registration method based on improved ICP are proposed.Then,in the surface reconstruction of point cloud data,an improved Poisson surface reconstruction method is proposed.Specific research contents are as follows:1.A complementary hybrid filtering strategy is proposed for point cloud denoising.Through the research and further improvement of the voxel filtering and statistical filtering in the point cloud filtering algorithm,the effective parameters of removing the outliers of the point cloud data are clearly defined.Then,an approximate voxel filtering method with minimum point constraint is proposed to dynamically adapt the algorithm to the sparse condition of each part of the point cloud.The voxel is approximately represented by the adjacent points of the center of gravity in the voxel.This method can effectively reduce multi-scale noise without damaging the geometric structure of source data.2.A combination of Intrinsic Shape Signatures(ISS),Three Dimensional Shape Context(3DSC),Random Sample Sonsensus(RANSAC)algorithm and Improved Iterative Closest Point(ICP)algorithm is proposed for point cloud registration.Firstly,voxel filter is used to subsample,ISS algorithm is used to extract the feature points,and 3DSC is used to describe the feature points.Then,the characteristics of ISS-3DSC and RANSAC algorithm were used for rough registration.Finally,ICP algorithm with direction vector threshold constraint is used for accurate registration.This algorithm can shorten the registration time effectively while maintaining high registration accuracy.3.An improved Poisson reconstruction algorithm based on boundary constraints is proposed for surface reconstruction.Firstly,octree was used to replace K-D tree for neighbor search.Then Open Multi-Processing(Open MP)was used to accelerate the normal estimation based on the moving least squares algorithm,and the minimum cost spanning tree was used to adjust the consistency of the normal direction.Finally,the shielding Poisson algorithm with Neumann boundary constraints was used to reconstruct the point cloud.This method can shorten the reconstruction time and reduce the pseudo-surface generation.4.According to the above research methods,a 3D reconstruction platform based on point cloud data processing is designed and implemented,which provides a convenient and concise visual interface.It can complete the operation of reading and saving,filtering,registration and surface reconstruction of point cloud data through human-computer interaction,and the function of each module is experimentally verified and expounded.