Polint Cloud Registration Based On Local Descriptors

The descriptors of the traditional 3D point cloud registration algorithm based on manual descriptors are low in robustness,and the generated matching pairs have only local correspondence,and there is no association of global information.In view of the limitations of traditional descriptors,a registration method based on PointNet Siamese network is proposed for point cloud local descriptors.The main contributions of this paper can be summarized as the following two aspects:First,a 3D point cloud descriptor based on PointNet Siamese network is constructed,which is more robust than traditional descriptors.Second,a method for calculating the similarity coefficient between matching pairs of point clouds is proposed.Based on the algorithm associated with community detection,the matching pairs of 3D point cloud point errors are eliminated,and the accuracy of registration is improved.The main research work of this paper is as follows:Point cloud local descriptor based on PointNet Siamese network.The PointNet network is a point cloud data-based network.The Siamese network is a neural network structure used for similarity discrimination in two-dimensional images.By combining the PointNet network with the Siamese network structure,the similarity of point cloud data is discriminated.The output of a single branch of the Siamese network is used as a local descriptor for point cloud data.The method is tested on two test scenes and compared with the traditional 3D point cloud descriptors.In this paper,the mean value of the error ? is 0.3358 and the mean of the variance is 0.1770,which is lower than the traditional method(FPFH,SHOT).The robustness of the descriptor proposed in this paper is verified.Elimination of incorrect point cloud matching pairs based on graph structure.Firstly,the similarity of the distance and the angle between the pair of point cloud matchling points is established,and the relationship between the point cloud matching pairs are established.Further,the graph structure and the similarity matrix are used to represent the global information of the point cloud matching pair.Then,based on the graph structure of the 3D point cloud matching pair,the correct matching pair of candidates is obtained by using the spectral clustering method,and the final correct matching pair is obtained according to the constraint of the same name point and position information.Finally,the SVD method is used to solve the rigid body transformation relationship,and the registration of the point cloud data is completed.The method is tested on three test scenes and compared with RANSAC and game theory.The proposed method is the best in terms of speed and precision,and verifies the effectiveness of the proposed method.