Abstract: Point cloud registration is a key technology in computer vision, three-dimensional spatial computing, and related fields. To address the insufficient point feature representation capability in partially overlapping scenarios, as well as the mismatching problems caused by noise interference and non-overlapping regions, we proposes a partial-overlap point cloud registration method that integrates dual-graph convolution and overlap-aware cross-attention. First, an enhanced edge-feature local neighborhood graph is constructed by integrating the center point, neighboring points, relative displacement, and distance scale. The neighbor-hood features are then fused through attention-based weighting to extract point features containing mul-ti-scale local geometric information. Subsequently, contextual features under cross-domain consistency constraints are constructed based on dual-graph convolution with a geometric graph and a feature graph, and are fused with point features to form cross-domain enhanced point features. Then, self-attention and cross-attention are employed to achieve intra-point-cloud and inter-point-cloud feature interaction, thereby obtaining the overlap score of each point and the point clouds in the overlapping regions. On this basis, an overlap-aware cross-attention module is designed, in which the overlap scores are introduced as priors into the cross-attention mechanism to strengthen the feature consistency of overlapping regions. Finally, to al-leviate the mutual interference between rotation and translation in joint estimation, a dual-branch parallel feature pathway for rotation and translation is constructed, through which soft correspondence matrices are obtained separately to solve the rigid transformation and improve the accuracy of pose estimation. Exper-imental results on the ModelNet40, 3DMatch, and KITTI datasets demonstrate that, compared with most mainstream methods, the proposed method achieves higher registration accuracy and better robustness, and can effectively realize stable registration of partially overlapping point clouds.