Research And Application Of Point Cloud Registration Based On Deep And Shallow Features Combined With Attention Enhancement

Point cloud registration refers to the process of calculating coordinate transformation relationship.The registration algorithm is used to integrate the point cloud data from different perspectives into the specified coordinate system through rigid transformations such as rotation and translation.Recently,the application of high-precision sensors has promoted the development of point cloud registration technology.Point cloud registration is the basis of three-dimensional surface reconstruction,and its registration accuracy directly affects the quality of three-dimensional reconstruction.At present,point cloud registration method based on deep learning is a hot research method.It’s hard to be disturbed by outliers and noise,and can effectively improve the precision of point cloud registration.This paper focuses on the point cloud registration method based on deep learning.Based on the full study of the existing advanced point cloud registration networks,a new point cloud registration network IADSPPCR is proposed.It is applied to partial point cloud registration to improve the precision of partial point cloud registration.Based on IADS-PPCR,a geometric feature enhanced network IADS-PPCR(PPF)is designed to further enhance the precision of partial point cloud registration.The work of this paper is mainly can be summarized in the following four aspects:(1)Studied the relevant knowledge of deep learning and convolution neural network.Focused on the relevant theories of point cloud registration,and numerically realized representative point cloud registration methods based on deep learning.The challenges of point cloud registration and its important value in practical application are pointed out.(2)A partial point cloud registration network IADS-PPCR is proposed.IADS-PPCR network is composed of deep and shallow features extraction,attention enhancement,prediction constraint parameters,feature matrix matching and singular value decomposition.The deep and shallow features extraction module provides effective and rich point cloud deep and shallow features for the attention enhancement module;The attention enhancement module selectively enhances or suppresses the deep and shallow features of the source and target point clouds;The prediction constraint parameters module adaptively matches the source and target point clouds in a learning way to obtain the association constraint parameters between the source and target point clouds;The feature matrix matching module fuses the enhanced deep and shallow features,constraint parameters and target point clouds to obtain the predicted target point clouds;The singular value decomposition module obtains the translation vector and rotation matrix between the source and the prediction target point clouds.Numerous experimental results show that this method can not only effectively improve the precision of partial point cloud registration,but also insensitive to noise and outliers.(3)The IADS-PPCR network is optimized,and a partial point cloud registration network IADSPPCR(PPF)is proposed.Compared with IADS-PPCR,which uses point coordinates and its domain point coordinates to describe point cloud features,IADS-PPCR(PPF)uses point coordinates,neighborhood point pair vectors and neighborhood point pair features to describe point cloud features.More geometric features are introduced to enhance the feature description ability of the network.The experimental results show that IADS-PPCR(PPF)method with rich geometric features can calculate and obtain high-precision point cloud transformation matrix.(4)Poisson and greedy projection triangulation algorithms are used to realize the threedimensional surface reconstruction of this paper method and the corresponding comparison method,and surface reconstruction results are compared and analyzed.