Research On Point Cloud Registration Algorithms Based On Quadric Error And Gaussian Mixture Model

3D point cloud processing technology is one of the important research topics of 3D modeling and computer vision.The technology is applied in every aspect of life,such as automatic driving,virtual reality,cultural relic restoration,robot visual positioning,etc.Among them,point cloud registration is an important point cloud processing technology.Since point cloud acquisition equipment can only obtain partial point cloud data from multiple single perspectives,it is necessary to use point cloud registration method to synthesize partial data into complete point cloud through spatial transformation.Point cloud registration is divided into global registration and local registration according to the relative position of point clouds.Starting from the method based on feature extraction and probability density function of global registration,this paper analyzed the shortcomings of the current registration method and optimizes the registration efficiency,accuracy and pose of the algorithm.The main work of this paper is as follows:(1)Research on point cloud registration algorithms based on quadratic error.In view of the high computational complexity of the current feature extraction algorithm,slow registration efficiency,and easy mismatch under large-angle registration,combined with the advantages of simple and stable geometric features extracted by quadratic error algorithm,two algorithms are proposed in this paper,one is the registration algorithm based on the global cost sequence of quadratic error,this algorithm uses all the geometric features of the point cloud to participate in the registration,and uses the grouping average method to make the registration time not prolonged with the increase of the number of points,so the efficiency is maintained.The second is the registration algorithm of strong feature point group based on quadratic error,which searches for neighboring strong feature descriptors and only uses strong feature descriptors to participate in registration,eliminating the errors caused by weak feature descriptors and improving the registration accuracy.The two proposed algorithms have shown good performance in the experiments,especially in the large-angle registration experiments,other registration algorithms have a certain degree of mismatch,and the algorithm in this paper still achieves good results.(2)Research on point cloud registration algorithm based on gaussian mixture model.At present,the commonly used coherent point drift algorithm has some problems such as high computational complexity,long algorithm iterations and unreasonable outlier processing.We propose three improvements,first,a permutohedral lattice filtering model was introduced to replace the probability density matrix in the original algorithm,which effectively speeds up the matrix calculation.Second,the global square iteration method is used to optimize the parameters of EM algorithm,and the regularization constraint is given to limit the jump amplitude,so that the result can reach the target value quickly and stably.Third,set the iterative outlier formula,which is used to correct the interference ability of noise and outliers in the iterative process,so as to improve the accuracy of the algorithm.Compared with other popular algorithms and the original algorithm,the improved coherent point drift algorithm has better performance than the original algorithm and exceeds other algorithms.