Abstract: Objectives: Significant initial deviations in the Exterior Orientation Parameters (EOPs) between aerial imagery and Airborne Light Detection and Ranging (LiDAR) data often lead to insufficient registration accuracy, adversely affecting the fusion of multi-source geospatial data. To overcome this challenge, a registration method is proposed that integrates both global distribution and local contour features of buildings. The primary goal is to substantially enhance the fusion accuracy of these two data types, thereby providing a reliable foundation for critical applications such as high-precision change detection, refined feature extraction, and high-fidelity 3D real-scene modeling. Methods: The methodology comprises three key stages. First, in the global coarse registration stage, Structure from Motion (SfM) technology is used to reconstruct a sparse 3D point cloud from overlapping aerial image sequences. Subsequently, through efficient filtering and Euclidean clustering algorithms, individual building instances are segmented and extracted from the image-derived sparse point cloud and the LiDAR point cloud, respectively. Given the relatively stable global layout characteristics of buildings in urban spaces, a deep learning-based Graph Matching Network (GMN) is introduced to abstract the scattered building instances into graph nodes. By learning and comparing the topological similarity of their spatial distribution relationships, cross-source building instance-level correspondences are established, thereby effectively correcting large-scale rotation and translation deviations. Second, in the local fine registration stage, we shift our focus to the geometric details of the buildings themselves. Building contour line segments are accurately extracted from aerial images and the LiDAR point cloud, respectively. On this basis, rigorous screening and matching are performed using multiple geometric constraints, including directional consistency, length similarity, and spatial proximity between line segments, to construct a set of high-confidence corresponding line segment pairs. Finally, based on the above reliable matching primitives, an overall optimization function is constructed with the objective of minimizing the cumulative spatial Euclidean distance between all corresponding line segment pairs. To enhance the robustness of the algorithm, the Random Sample Consensus (RANSAC) algorithm is integrated during the optimization process to solve for the optimal aerial image EOPs. This iterative approach identifies and eliminates potential mismatches, ensuring the accuracy and stability of the solution. Results: To validate the effectiveness and universality of the proposed method, experiments were conducted on two datasets with different point cloud densities and different image resolutions. Quantitative analysis results demonstrate that after registration using the proposed method, the average projection errors on the two datasets were significantly reduced to 0.26 times and 0.5 times the average point spacing of their corresponding LiDAR point clouds, respectively. This accuracy significantly outperforms traditional registration methods based on single features, proving the excellent performance of our method across different data specifications. Conclusions: The synergistic optimization mechanism integrating global distribution constraints and local line feature matching effectively overcomes modal transformation errors and initial extrinsic parameter deviations, establishing a high-precision registration foundation for multi-source geospatial data fusion.