Abstract: A network called RPMNet++ has been proposed to improve the accuracy of point cloud registration in complex scenes and under non-ideal sample conditions, including noise interference, density inconsistency, and structur-al incompleteness or damage caused by occlusion. First, we construct the Copula denoising module. Based on the assumption that neighboring data points have similar features, we extract point cloud features using a con-volutional neural network. These features are then used to calculate the Kendall correlation coefficient τ and the Clayton Copula distribution function. Based on the calculation results, positively correlated inliers are retained while negatively correlated noise points are filtered out, thereby mitigating issues such as feature bias, parame-ter estimation errors, and misidentification of corresponding points due to noise interference. Second, local fea-ture learning is achieved under the bidirectional attention mechanism. Bidirectional attention is explicitly di-vided into two parts: attention from sampling points to neighboring points and attention from neighboring points to the central point. By combining these two components and incorporating neighborhood feature encoding, the learning of sampling point features and their spatial correlations in the neighborhood is enhanced. This facili-tates the effective extraction of data features from the denoised, sparse, and locally incomplete point cloud. This feature learning method ensures the integrity of neighborhood correlation features while improving the net-work’s ability to learn fine-grained local features of point cloud data. Point cloud registration experiments demonstrate that, compared to RPM-Net on the ModelNet40 dataset, RPMNet++ significantly reduces the iso-tropic average rotation error and translation error, with reductions of (0.026, 0.001), (0.267 0, 0.001 9), and (0.560, 0.007) on noise-free, noisy, and partially visible noisy datasets, respectively. On the Stanford 3D dataset, RPMNet++ outperforms seven cross-source registration algorithms, achieving optimal results and exhibiting good generalization performance and application value.