To address the challenges caused by variations in point cloud quality during cross-source registration, we introduce an improved local shape descriptor, termed the Spherical Voxel Center Descriptor (SVCD). This descriptor forms the foundation of a robust method designed to mitigate the effects of density and distribution disparities in cross-source point clouds. The core innovation lies in the design of a dual-weighted Local Reference Frame (LRF) computation combined with spherical voxel segmentation of local surfaces, effectively mitigating the impact of density and distribution discrepancies in cross-source data. SVCD enhances feature descriptiveness and robustness by encoding the distances from voxel centers to keypoints. Accurate correspondences are established using nearest neighbor similarity ratios, and the final rigid transformation matrix is derived via singular value decomposition. Extensive experiments on 3DCSR and real-world datasets demonstrate the superior performance of SVCD, achieving a registration error as low as 0.0048. Recall rates of 82.83% and 83.45% represent improvements of 10.24 and 11.16 percentage points, respectively, over existing methods. Additionally, SVCD exhibits exceptional robustness to Gaussian noise, attaining an average recall of 76.54%, significantly outperforming other descriptors. These results underscore the effectiveness and reliability of SVCD for cross-source point cloud registration.