Research Of Point Cloud Registration Technology Based On Feature Fusion

Point cloud registration is one of the key research problems in computer vision,which purpose is to obtain a transformation matrix to make source point cloud and template point cloud perfectly coincide.The closer the value of the transformation matrix is to the ground truth,the better the registration effect is.Point cloud registration technology is widely used in 3d reconstruction,3D positioning,pose estimation and other computer vision fields.This paper focuses on point cloud feature extraction,feature matching and other methods.Meanwhile,studies 3D point cloud registration by combining existing 3D point cloud processing network and dynamic graph edge convolution network technologies.The main contents include:Existing feature extraction and registration methods of 3D point cloud are studied in depth in this paper.In view of the existing terminal to the end to end point cloud registration method due to insufficiency of feature extraction and matching problem of low precision and the fusion of global and local features PTRNet point cloud registration model was proposed,which uses the original point cloud as input,global feature extraction of point cloud based on attention technology first,then edge convolution module is used to encode the local characteristics of point cloud,By integrating it with global features,a more comprehensive global representation can be obtained.Finally,multilayer perceptron is used to generate transformation matrix.Comparative experiments were carried out on public data sets,and the results were evaluated by the mean square error and root mean square error.The results show that compared with the existing non-corresponding registration method and the point cloud registration method based on optimization,the PTRNet method has the characteristics of smaller error,stronger robustness and higher registration accuracy.The existing depth registration methods which directly extract the features of the input point cloud often have low precision due to ignoring the geometric features of the point cloud.Structuring the point cloud image is easier to extract the subsequent point cloud features,and the correlation information between two point clouds can be better extracted by obtaining the geometric features of point clouds.Inspired by graph attention network and combined with dynamic graph convolution network,the geometric and correlation information features of point cloud were extracted,and CPCRNet was proposed for deep registration.In order to obtain multi-level and more semantic representation of point cloud,dynamic graph edge convolution neural network is used to extract feature from point cloud data firstly.And a new feature fusion module is proposed to obtain the correlation information between point clouds by structuring the feature graph of point cloud.Finally,singular value decomposition is used to generate transformation matrix.A large number of experiments have been carried out on the public data set and the local part data set.The experimental results show that CPCRNet method not only has the characteristics of high registration accuracy and strong robustness,but also can be applied to the actual data and has universality.