Research And Application Of Point Cloud Registration Based On Deep Learning

Point cloud registration is widely used as a key technology in 3D reconstruction,target recognition,and camera relocation.In practice,although the ICP(Iterative Closest Points)algorithm can achieve high registration precision,the traditional initial registration algorithm has a low success rate.This instability limits the use of registration algorithms in industry.Aiming at the problem of unstable registration in project,the deep learning method is used to research the feature-based initial registration algorithm.First,Point Net is used to train local features suitable for point cloud registration,which improves the recall and precision of the correspondences in the initial registration.Then based on the high recall and high precision of the Point Net feature,the registration algorithm is optimized by using iterative correspondences,which improves the registration success rate of the registration algorithm.Finally,the improved algorithm is successfully applied to the complex parts measurement and the 6-DOF pose estimation.The main contents include:(1)Research on the features used for point cloud registration.Aiming at the problem of the difference in the arrangement of point clouds obtained from different point clouds,the local point cloud features based on Point Net are designed and trained by using the Scene NN dataset and the Triplet algorithm.And the Point Net feature is superior to traditional point cloud feature in terms of recall and precision.(2)Research on registration algorithms based on the Point Net feature.Combined with the Point Net feature,the initial pose of registration is solved by using SVD,and the corresponding points' distances are calculated by using the initial pose.Then sort the correspondences by the distances and delete the correspondences with large distances.Finally,iterate it to improve the registration success rate.(3)Apply the registration algorithm to the actual scene.By registering the part to be measured of the point cloud model of complex parts and segmenting the mark point cloud for measurement,the complex part measurement application is implemented;in addition,by using the pre-selected measured template to register the object,the transformation matrix is obtained and the 6 DOF pose estimation application is implemented.In general,this paper uses deep learning to improve the initial registration success rate,and achieves stable measurement of complex parts and 6 DOF pose estimation applications.