In the registration of point clouds with low overlap rate, traditional methods often struggle with features and matching difficulties, frequently falling into local optimum under large pose errors or complex transformation scenarios, which compromises registration accuracy. To solve these problems, an adaptive graph convolution model with progressive feature fusion pyramid network is designed to establish correspondences between point clouds from coarse to fine scales. Firstly, Adaptive Graph Convolution (AGConv) is employed to extract and encode spatial features, and then Progressive Feature Pyramid Network (AFPN) is utilized to fuse semantic information across multiple scales to jointly enhance the performance of the model in complex 3D scene understanding and analysis tasks. Secondly, a geometric Transformer is introduced to strengthen the model's comprehension of global structures and correlations, and achieve high-quality super-point matching. Finally, a local-to-global registration method is designed by integrating AGConv and AFPN, leveraging learned local point features from the backbone network and resolving global ambiguity problems through superposition point matching, thereby improving the robustness of the algorithm. Experiments show that the proposed network significantly improves the registration accuracy of point clouds with low overlap rate.