Local feature extraction network with high correspondences for 3d point cloud registration

3D point cloud registration is an important task in computer vision. Due to the irregularity of point clouds, it is still a challenging problem to realize the accurate registration. Recently, with the development of deep learning, scholars have proposed many learning-based methods, which can enhance the correspondence of points and not rely on the initial alignment conditions. However, most works tend to ignore the importance of local features, leading to the unreasonable matching. To solve this issue, we propose two networks to extract richer local information. In order to find a closer internal relation between the points, a Subtract Attention Network (SANet) is designed. In which, we propose a Subtract Attention Module (SAM) to aggregate the point-wise feature representations and construct the key points of feature space on this basis. We also propose a Position Encoding Network (PENet) to determine the spatial correlation with the utility of local coordinates. After combining the spatial features of different dimensions, the connections of key points in the feature space tend to be more credible. Thus, we can effectively obtain the local correspondence between each point and then improve the accuracy of registration. The results on the commonly used dataset ModelNet40 show the superiority of our method.