An attention guided local global feature fusion network for robust point cloud registration

Abstract Point cloud registration in complex environments remains challenging because measurement noise, partial overlap, and structural variations can reduce correspondence reliability and registration robustness, while computational efficiency is also critical for practical deployment. These difficulties are closely related to insufficient adaptive modeling of local geometric structures, inadequate suppression of weak global graph connections, and limited exploitation of inter-cloud similarity and discrepancy information. To address these issues, we propose AGLGNet, an attention guided local global feature fusion network for robust point cloud registration. First, the Adaptive Local Neighborhood Graph Calibration Module constructs geometry-guided neighborhood representations and adaptively recalibrates neighboring features, thereby enhancing local structural discrimination and reducing sensitivity to noise. Second, the Dynamic Weight Global Graph Learning Module adjusts graph connection weights according to inter-point correlations, suppressing weak relationships and strengthening long-range contextual modeling. Third, the Hybrid Attention Feature Fusion Module jointly models inter-cloud similarity and discrepancy information through cross-attention and further refines the fused representations using self-attention. Extensive experiments on 3DMatch, ModelNet40, KITTI, the Stanford 3D Scanning Repository, and real-world workpiece data demonstrate that the proposed method achieves low registration errors and strong robustness under complete, partial, and noisy point cloud conditions. It also maintains an inference time on the order of 10 −2 s, providing a favorable balance among registration accuracy, robustness, and computational efficiency.