Research On 3D Point Cloud Registration Algorithm Based On Quadric Error And Gaussian Mixture Model

In the field of computer vision,with the continuous development of 3D scanning equipment and 3D processing capabilities,computer vision has attracted more and more attention.As one of the most basic and important parts of machine vision,3D point cloud registration is widely used in all aspects of modern life,including mechanical engineering,medical imaging,virtual reality,augmented reality,robot vision,etc.Since the 3D scanning device cannot obtain all point cloud images at once,it is necessary to align the point clouds obtained from different directions into a common coordinate system to obtain a complete three-dimensional model.Three-dimensional point cloud registration can be divided into coarse registration and fine registration according to the degree of registration refinement.This paper mainly focuses on the feature-based registration method in the coarse registration and the probability-based method in the fine registration,and analyzes the shortcomings and defects of the existing algorithms,from the optimization of registration accuracy,efficiency of the registration,and registration attitude.The main work of this paper is summarized as follows:1.Research on point cloud registration based on Gaussian mixture model of filtering.First,aiming at the problem of high registration complexity of high-dimensional Gaussian mixture model,this paper transforms the calculation of high-dimensional Gaussian mixture into the process of lattice filtering through permutohedral lattice,and reduces the computational complexity of Gaussian mixture to linear magnitude.Second,for the problem that the point-to-point distance parameter cannot capture the local features of the plane,the point-to-surface distance parameter is introduced,which can more accurately determine the corresponding point,accelerate the convergence speed of the registration algorithm,and further improve the registration efficiency of the algorithm.Finally,experiments are carried out on the standford data set,and results show proposed algorithm improves the registration accuracy and speeds up the registration process.2.Research on point cloud registration algorithm based on quadratic error.Frist,this research introduces the concept of quadratic error into point cloud registration,and combines the excellent characteristics of quadratic error in model simplification,and proposes a feature description method based on quadratic error.In our coarse registration method based on quadratic error,the quadratic error of points is taken as the feature description of the registration algorithm,which is superior to other popular algorithms in large-scale rotation.Second,In the optimization method of coarse registration,the cost matrix of each neighborhood point is obtained by calculating the quadratic error of each neighborhood point,which improves the registration accuracy.In the fine registration algorithm,the quadratic error is used as a local feature in the registration framework based on probabilistic model,which not only keeps the robustness of probabilistic model to noise and outliers,but also reduces the sensitivity of probabilistic model to large-scale rotation.Third,experiments were carried out on the standford data set.In the coarse registration experiment,the registration algorithm based on quadratic error is superior to most feature-based algorithms in terms of accuracy,efficiency and larger-scale rotation.in the experiment of fine registration,The Gaussian mixture model algorithm based on quadratic error is obviously better than other registration algorithms based on Gaussian mixture model in large-scale rotation.Finally,summarize the content of the thesis,think about the deficiencies in the thesis,and prospect for further research.