The disordered and unstructured characteristics of 3D point cloud data pose challenges to existing point cloud registration algorithms.Based on this

a graph convolutional point cloud registration network that fuses multi-scale features is proposed aiming at existing registration algorithms.This network uses dynamic graph convolutional neural network(DGCNN)as the backbone network.K-nearest neighbor(KNN)is introduced in edge convolution to extract the point pair features(PPF)of the center point and its neighboring points

and relative spatial positions between the point pairs are encoded to enhance the structural relationships of the point cloud.In addition

the feature expression ability of graph convolution is improved through skip connections between adjacent edge convolution layers.Then

the correlations between different points of the point cloud are captured through the Transformer and Pointer network modules

and the relative pose relationship between the two point clouds is estimated.The rotation matrix and translation matrix of the rigid transformation are extracted through the differentiable singular value decomposition(SVD)layer.Finally

according to the changes of the point cloud features extracted during the registration process

the weight values of each component of the loss function are dynamically adjusted to obtain more accurate model prediction results.The proposed method is validated on the public datasets ModelNet40 and ShapeNet.Compared with existing advanced point cloud registration methods

the proposed method has the advantage of low registration error.