Abstract: Point cloud registration is a key issue in the field of 3D computer vision. To address the limitations of tradi-tional point cloud registration methods in the absence of surface normal constraints and in the registration of partially overlapping point clouds, a novel method for partially overlapping point cloud registration without surface normal constraints is proposed. This model first employs a probabilistic overlapping infer-ence mechanism to obtain the corresponding overlapping subsets and eliminate the influence of non-overlapping regions. Then, a layer-by-layer stacked feature extraction network is constructed to grad-ually expand the receptive field and enhance the geometric feature expression ability. Next, a unidirection-al self-cross attention mechanism is adopted to enhance the interaction between features. Finally, singular value decomposition is performed on the feature similarity matrix to achieve end-to-end rigid transfor-mation parameter estimation, solving the problem of registration without vector constraints. Training and validation were conducted on the ModelNet40 dataset, and the experimental results were compared with those of mainstream point cloud registration methods. The results show that the MSE(R) value is 18.079, the RMSE(R) value is 4.252, the MAE(R) value is 2.179, the MSE(t) value is 0.002, the RMSE(t) value is 0.047, and the MAE(t) value is 0.033. This provides a new solution and research direction for the registra-tion of partially overlapping point clouds.