This paper addresses the challenges of inefficient registration in point cloud data processing due to volume and complexity of spatial structures by introducing an optimized point cloud registration algorithm. The algorithm integrates a refined point cloud with an initial rough registration step. Key points are identified through voxel sampling, leveraging the variance in mean angles between neighborhood points and the center point normal line. The Fast Point Feature Histogram (FPFH) serves as the feature descriptor. In the search of correspondence relationship, according to the similarity of vector angle between adjacent matching pairs and Random Sampling Consistency (RANSAC) algorithm, filtering optimization is carried out to accurately correspond to the relationship and complete rough registration. Finally, the ICP algorithm is used to achieve accurate registration. The results show that the proposed algorithm effectively captures key points in regions of significant spatial variation, providing an advantageous initial position for precise registration and significantly reducing overall point cloud registration time.