RGNN-M: Residual-Enhanced Graph Neural Network with Multi-Geometric Constraints for Point Cloud Registration

To address issues such as deep feature propagation loss and insufficient geometric constraints in existing point cloud registration algorithms, a point cloud registration method based on residual-enhanced GNN and multi-constraint fusion is proposes on this paper. This method extracts hierarchical features using a residual-enhanced GNN. It incorporates channel attention to reinforce key information and integrates high-dimensional features with geometric attributes. A dual-branch network is employed to screen key feature points, followed by constructing a multi-scale feature matrix for similarity calculation. Attention weighting is used to optimize matching weights, and finally, weighted SVD is applied to solve the pose. Additionally, a multi-constraint loss function is designed to optimize the model. The experimental results demonstrate that the proposed algorithm outperforms the comparison algorithms in terms of registration accuracy and robustness under unseen categories, unseen shapes, and noisy scenes. The RMSE(R) can be as low as 0.68, with the lowest RMSE(t) reaching 0.005. Additionally, the algorithm maintains stable performance even under noisy conditions, fully validating the effectiveness of the proposed method.