Abstract: Objectives: Point cloud registration is a core fundamental technique for tasks such as indoor 3D reconstruction and indoor localization. Indoor scenes generally suffer from weak textures and repetitive structures, which easily lead to insufficient discriminability of point cloud features. It is difficult to construct stable and reliable corresponding point relationships, severely limiting the accuracy of point cloud registration. To address this problem, we propose a geometry-aware enhanced coarse-to-fine registration method for indoor point clouds. Methods: In the coarse-scale stage, a multi-scale feature aggregation method with spatial information embedding is first introduced to enhance the discriminative ability of coarse-grained superpoint features. On this basis, a superpoint overlap prediction module is designed to filter non-overlapping superpoints by generating overlap masks, which effectively suppress the interference of invalid matching. Meanwhile, position prior information is integrated for superpoint matching, and geometric position combined with feature representation is modeled to improve the reliability of superpoint correspondence. In the fine-scale stage, fine-grained point features, together with the point-to-node grouping strategy and Sinkhorn algorithm, are adopted to refine superpoint matching results and solve high-precision point-to-point correspondences. Finally, the random sample consensus algorithm is used to estimate the rigid transformation parameters between point clouds. In the training phase, a composite loss function consisting of superpoint matching loss, overlap prediction loss and point matching loss is constructed to realize collaborative convergence optimization of multiple modules. Results: Qualitative and quantitative experiments were conducted on the publicly available indoor 3DMatch and low-overlap 3DLoMatch datasets. The results show that the proposed method achieves state-of-the-art performance on the 3DMatch dataset, with RTE, RRE, and RR reaching 0.061 m, 1.93°, and 90.4%, respectively. Compared with existing methods, the rotation and translation errors are reduced by 4.7% and 3.5%, respectively. On 3DLoMatch, the proposed method achieves an RTE of 0.091 m, an RRE of 3.14°, and an RR of 63.9%, demonstrating good registration capability in low-overlap scenarios. Conclusions: The proposed method effectively improves the accuracy and robustness of point cloud registration in complex indoor scenes.