The Research Of Point Cloud Image Registration Technology Based On Deep Learning

Point cloud registration is an important basis of 3D reconstruction and is widely used in many fields.The quality of point cloud registration algorithm is related to the speed and accuracy of 3D reconstruction.How to register point cloud images more quickly and accurately is a hot and difficult point in the research of 3D reconstruction technology.At the same time,the processing speed and registration accuracy of the traditional point cloud registration algorithm are no longer suitable for the demand.The point cloud registration algorithm based on deep learning is proposed one by one.Moreover,the point cloud features extracted based on deep learning show better effect than the non learning algorithm.The registration method based on deep learning also effectively improves the registration accuracy and speed.In this paper,the point cloud image registration technology is deeply studied in order to design a point cloud registration algorithm based on deep learning,and use lightweight neural network to ensure the registration speed and improve the registration accuracy at the same time.The main research work of this paper is as follows:This paper proposes an adjustment network based on PCA.The network module can adjust the location of the point cloud to be registered and save the computing cost of the whole network.Based on multi-layer perceptron and IC algorithm,this paper proposes a point cloud registration algorithm PDC net based on global feature extraction and deep iterative neural network.By maximizing the pool,the algorithm combines the feature maps of each MLP layer to form global features,and completes the registration in the high-dimensional feature space.The algorithm uses global features for feature matching without searching the corresponding point pairs for the two point clouds to be registered,so as to save time cost and hardware overhead,and the registration results are also constrained by the relative error between the global feature maps.When realizing point cloud image registration,the registration error does not increase with the increase of the initial position angle between the two point clouds to be registered,which is greatly improved compared with other methods.Moreover,the network avoids the algorithm from falling into the local optimal solution and enhances the robustness of registration.Considering the shortcomings of PDC net,this paper proposes a deep learning point cloud registration method KCP net based on local feature extraction and SVD,which improves the accuracy of point cloud registration and ensures the speed of point cloud registration based on PDC net algorithm.The algorithm extracts the local features of the point cloud by combining KC algorithm and MLP,generates a new corresponding point cloud by using the probability method through the local features,retains more information of the point cloud,and solves the transformation matrix by singular value decomposition method,which replaces the iterative cycle process,saves the calculation cost,improves the accuracy of point cloud registration and ensures the speed of registration based on many aspects.KCP net has good generalization and robustness,and high registration accuracy and efficiency.The MSE error of rotation angle reaches 10-2,the MSE error of translation distance is zero,and the calculation efficiency is stable within 100 ms.