A new method for point cloud registration: Adaptive relation-oriented convolution and recurrent correspondence-walk

For point cloud registration (PCR), a matching matrix is critical. Unfortunately, the existing approaches do not explicitly devise schemes to refine the matching matrix. Furthermore, previous studies focused on the design of feature interactions between two point clouds and lacked attention to the discriminative features required for point cloud registration. This study presents a novel PCR method called RecARONet. The method mainly includes two innovations: an adaptive relation-oriented convolution (ARO-Conv) with effectiveness and a recurrent refinement technique of the correspondence based on the adaptive neighbourhood consensus constraint, mainly for more accurate registration. Specifically, ARO-Conv reconstructs the node representation by weighting the relations in the local neighbourhood rather than generating point features from the embeddings of the neighbours. This simple but effective operation can reduce feature redundancy and alleviate structural smoothness to a certain extent. It can also assign appropriate weights to the relations and different channels of features to capture more distinct local topological information. In addition, a recurrent correspondence-walk with a semantic adornment algorithm based on the adaptive neighbourhood consensus constraint is depicted, which can adaptively capture the differences in the local structure among proxy point pairs and recurrently update correspondences. Registration evaluations were performed on several complete/partial point cloud datasets, which revealed that the constructed model achieved excellent performance.