Adaptive spatial feature extraction and graphical feature awareness for robust point cloud registration

In recent years Transformers have achieved significant success in the field of 3D vision due to their inherent advantages in capturing global correlations between features. However, this can be a drawback in point cloud registration, especially in scenes with low overlap rates, where a large number of non-overlapping points can lead to ineffective or even negative attention allocation. Moreover, existing RANSAC-based registration estimators usually require a large number of iterations to obtain acceptable results, resulting in significant computational overhead. To address the above issues, we propose LDGR, which achieves robust registration in low overlap scenarios by utilizing a feature extractor with adaptive receptive fields and graphical feature awareness. Firstly, we proposed a 3D convolutional method with an adaptive receptive field named Adaptive Point Convolution (APConv) as the feature extractor. Its distinguishing feature is that the receptive field of the convolutional kernel is obtained through learning, which enables it to more flexibly handle irregular and unordered point clouds, thereby extracting richer and more diverse point features. Furthermore, to overcome the dilemma in cases of low overlap, we improved the transformer with rich local geometric information embedding and graphical feature awareness. This ensures that the model focuses more on the local spatial structure and features of the points during low overlap registration. Additionally, we propose a registration evaluator with local diffusion to global (LDGR). Compared to traditional RANSAC, it achieves comparable registration quality without requiring numerous iterative computations. Finally, we conducted several experiments on publicly available datasets such as 3DMatch and 3DLoMatch, KITTI odometry, ModelNet and ModelLoNet to validate the effectiveness of our method. We achieve optimal results in all four tests on ModelNet and ModelLoNet, significantly outperforming current state-of-the-art methods. Results on the challenging 3DMatch and 3DLoMatch datasets demonstrate the robustness of our method, with our inlier ratio substantially outperforming current state-of-the-art methods. Our experiments on the KITTI dataset demonstrate that LDGR performs no worse than RANSAC, while not requiring a large number of iterations.