Research On 3D Point Cloud Registration Method Based On Deep Learning

Recent years,deep learning are becoming a hot research among academic institutions and scholars.As a result,the research on deep neural network based point cloud classification and segmentation has made a significant breakthrough.Scholars in various fields took the advantages of deep learning in feature extraction and large amount of data processing to solve specific problems,one of which is deploying point deep learning method to address the cloud registration.In this paper,on the basis of related researches,two point cloud registration methods that are all based on the deep learning are proposed.By taking advantages of deep learning network in feature extraction,we propose a point cloud registration algorithm that is based on dynamic graph neural network.Our algorithm deploys the dynamic graph convolution neural network(DGCNN)to extract local features of the point cloud;then the point with prominent features will be found as the key point by using point weighting;next,the soft connection between the key point sets is estimated by using the similarity of features;and the final rigid transformation relationship is calculated by using singular value decomposition method.When implementing,we integrate each module as an end-to-end network for training,making each module in the front serve the final registration results,which improves the accuracy and generalization of the point cloud registration,moreover,the registration speed is faster.Based on the robust point matching algorithm,we propose a deep learning based RPM algorithm.Our algorithm not only saves the robustness of the robust point matching algorithm to noise points and outliers,but also improves the high sensitivity to the initial positions.The proposed algorithm extract the mixed features of point cloud by combining local feature descriptor construction and feature extraction network,then the mixed features are used to replace the position information on estimating the corresponding relationship between point clouds,overcoming the limitations that the original algorithm is sensitive to the initial positions.Furthermore,a multi-layer perceptron(MLP)is used to estimate annealing parameters according to the current registration situation,solving the problem of manual parameter adjustment At last,to make full use of the learning ability of the deep learning network,we use an end-to-end training mode to implement the whole training processing,improving the noise robustness and registration effect of some missing point clouds.Finally,the two methods are implemented in 64 bit Ubuntu 18.04,python3.7,Py Torch1.3.1.Compared with other previous registration algorithms,we demonstrated and analyzed the registration accuracy,generalization,noise robustness and registration time of our two algorithms.