Point Cloud Registration Network Based on Convolution Fusion and Attention Mechanism

In 3D vision, point cloud registration remains a major challenge, especially in end-to-end deep learning, where low-quality point pairs will directly lead to the degradation of registration accuracy. Therefore, we propose a point cloud registration network based on convolution fusion and a new attention mechanism to obtain high-quality point pairs and improve the accuracy of registration. In this work, we first fuse kernel point convolution and adaptive point convolution by cross-attention mechanism as the feature extraction backbone of the network to obtain features. Secondly, we use transformer to exchange information between source and target point clouds, which consists of a new attention mechanism module, named ReSE-Attention. It obtains a global feature view by adding a squeeze extraction module and deep learnable parameters to the normal attention mechanism. And then, a regression decoder is adapted to generate the correct point pairs. Finally, we first introduce Focal Loss on the loss function in point cloud registration to balance the relationship between overlapping and non-overlapping regions. Our approach is evaluated on both the scene dataset 3DMatch and the object dataset ModelNet and achieves state-of-the-art performance.