Research And Application Of 3D Common-source Point Cloud Registration Technolog

As the most traditional 3D point cloud data registration algorithm,the registration efficiency of iteration closest point depends on the good initial pose of the point cloud.How to make the 3D point cloud data have a good initial pose before the iterative nearest point registration is a problem of great research value,starting from this problem,mainly including the following three aspects of work.(1)In order to solve the problem that the traditional ICP registration algorithm is more dependent on the initial pose,a point cloud registration algorithm based on improved fast point feature histogram and double iteration is proposed.Voxel filtering is used to downsample the 3D point cloud data,and the center of gravity of all point clouds in the voxel raster is saved for subsequent registration operations.Then,the normal vector calculated by the improved ISS algorithm is used,and the angle of the normal vector is used as a constraint to extract the feature points.Then,the improved FPFH algorithm is used to obtain the feature descriptor,and the average value and variance of the angle between the feature descriptor and the normal vector are used as constraints to obtain the matching pair.Then,the SVD algorithm is used to complete the coarse registration,and finally the double-iterative ICP algorithm is used to complete the fine registration.According to the analysis of experimental results,it can be seen that the proposed three-dimensional point cloud registration method can make the registration point cloud have a good initial posture,and the efficiency and accuracy are significantly improved.(2)In order to improve the efficiency and accuracy of low-overlap 3D point cloud and multi-noise 3D point cloud registration,a point cloud registration algorithm based on eigenvalue deviation ratio is proposed based on the structural characteristics of 3D point clouds.Firstly,the decentralized 3D point cloud is filtered by voxel filtering to remove noise,and then the deviation ratio of the eigenvalues of the covariance of each neighborhood point is used to obtain the feature points,and then the 3D point cloud spatial pose relationship is used for coarse registration,and finally the improved ICP algorithm of Kd-tree is used to complete the fine registration.Experimental results show that the proposed registration method has higher efficiency and accuracy than the registration method in the face of low overlap rate and multi-noise point cloud registration.(3)Apply the registration method proposed in this paper to practical applications,and use three-dimensional cameras,computers and robotic arms to complete the collection,registration,classification and final sorting of target point cloud data.