To address the registration challenges caused by cross-source point cloud quality disparities， this paper proposes an improved spherical voxel local shape descriptor （Spherical Voxel Center Descriptor， SVCD） for cross-source point cloud registration. SVCD effectively mitigates density and distribution variations through dual-weighted Local Reference Frame （LRF） computation and spherical voxel segmentation. Its core innovation lies in feature encoding based on the distance from voxel centers to keypoints， enhancing the distinctiveness and robustness of the descriptor. The registration process establishes correspondences via the nearest neighbor similarity ratio and solves the rigid transformation using the singular value decomposition. Experimental results on the 3DCSR and real-world datasets demonstrate that SVCD achieves a registration error as low as 0.004 8， with recall rates of 82.83% and 83.45% （improving baseline performance by 10.24 and 11.16 percentage points， respectively）， and the F1-scores are the highest （0.803 and 0.832）. In Gaussian noise experiments， SVCD maintains an average recall rate of 76.54%， significantly outperforming comparative methods， validating its strong robustness in complex scenarios. This method provides an effective solution for high-precision cross-source point cloud registration.