Point cloud registration plays an important role in 3D reconstruction and other point cloud-related tasks. The establishment of reliable and high-quality point correspondences is essential for accurately recovering the transformation matrix between point clouds. In recent years, the coarse-to-fine strategy have gained widespread attention to construct reliable point correspondences. However, in the coarse-scale superpoint matching stage, superpoints in non-overlapping regions can degrade the matching quality, thereby limiting the reliability of the refined point correspondences. To address this issue, this paper proposes a coarse-to-fine point cloud registration method based on superpoint overlap prediction, which focuses on optimizing the construction of superpoint correspondences at the coarse scale and effectively improving registration accuracy. Firstly, we employ a position-aware attention mechanism to enhance superpoint features under geometric constraints. Then, the superpoint overlap prediction module generates overlap masks based on the enhanced features, effectively filtering out superpoints over non-overlapping regions. This ensures that only the available superpoints over overlapping regions participate in the matching process, leading to more accurate superpoint correspondences and improved registration accuracy and robustness. Experimental results on indoor 3DMatch and 3DLoMatch datasets, as well as the outdoor KITTI dataset, demonstrate that our proposed method achieves the superior registration performance.