Research On Statistical Registration Algorithm Of 3D Laser Point Cloud

As a kind of sensing technology which can directly obtain the three-dimensional(3D)information of the target surface,3D laser scanning has unique technical advantages in the aspects of detection,recognition and reconstruction of the target under complex background.Point cloud registration,as an important part of point cloud processing,has developed into a hot research field in recent years,and many algorithms have been proposed by scholars at home and abroad.However,the point cloud data acquired by sensors such as Lidar and Kinect usually have such defects as large amount of data,large amount of noise,incomplete occlusion,size reduction and so on,which lead to the existing registration methods have some disadvantages,such as low registration efficiency and poor accuracy,so that it is difficult to meet the requirements for practical applications.Therefore,this paper mainly studies the point cloud registration algorithm under the conditions of large data volume,disorder,occlusion,missing,noise interference and the existence of scaling.Most of the existing 3D point cloud registration algorithms are based on statistics,including independent component analysis,Gaussian kernel function,principal component analysis and so on.In this paper,according to the correlation of mathematical statistics,a point cloud registration algorithm based on Kernel Canonical Correlation Analysis(KCCA)is proposed.This algorithm is based on the method of calculating the correlation of multiple sets of variables from the same object in statistics,and the goal is to maximize the correlation coefficient,so as to obtain the point cloud rigid transformation relationship.Firstly,the FPFH algorithm is used to search for the corresponding point cloud of the target point cloud in the source point cloud,the purpose is to make the geometric shape of the source point cloud and the target point cloud as consistent as possible.Then,the kernel canonical correlation analysis method is used to estimate the transformation matrix of the source point cloud and the target point cloud,and the rotation matrix is obtained on the basis of the transformation matrix,and then the translation vector is obtained.Finally,using open source data and field scan data to compare with the registration results of several traditional algorithms under different conditions,analyzes the advantages anddisadvantages of the algorithm under various conditions.Experiments show that the proposed algorithm has improved significantly in efficiency and accuracy.At the same time,according to the probability distribution characteristics of point cloud data itself,another point cloud registration algorithm based on Cauchy Mixed Model(CMM)is proposed.The algorithm does not consider the corresponding attributes of the two point cloud data,but only needs to solve the geometric relationship of rigid transformation on the basis of the probability distribution of the point cloud data itself.Since the order of points,noise and random absence of points do not change the probability distribution of the overall data,the algorithm has strong robustness.First,the Cauchy mixed model of the same order is used to fit the probability distribution of the source point cloud and target point cloud data respectively,and the point cloud registration model under rigid transformation is extended to the Cauchy mixed model.The maximum likelihood function is constructed based on the Bayes formula and Jason inequality,and the parameters in the mixed model are updated until convergence by the Expectation Maximization(EM)algorithm.Then,the rotation matrix is obtained by using the covariance matrix of the model corresponding to the maximum weight,then the translation vector is obtained,and the affine ratio is estimated in accordance of the center point of the corresponding model.Finally,using open source data and field scan data to compare with the registration results of several traditional algorithms under different conditions,highlights the strengths of the algorithm and analyzes its weaknesses.Experiments results show that the algorithm not only improves the accuracy but also improves the speed.At the same time,the algorithm can effectively register the point cloud in the case of affine and has good anti-noise ability.Finally,the two new algorithms are applied to 3D reconstruction based on point cloud registration,and three groups of local building scanning point clouds selected from point cloud Library(PCL)of Leibniz university are used for reconstruction,and good reconstruction results have been achieved.