To address the issues of redundancy in point cloud data, prone to mis-matched point pairs, and low alignment accuracy during the process of point cloud registration, a method that integrating supervoxels and geometric features is proposed in this paper. Firstly, key points are extracted using a combination of supervoxels and normal vector information. Subsequently, during the coarse registration phase, feature descriptions are generated using the Fast Point Feature Histograms (FPFH) method, and then initial correspondences are established based on the feature description using a bidirectional nearest neighbor ratio approach, and the correspondences are optimized using a normal vector angle strategy and the Random Sample Consensus (RASAC) algorithm to acquire a robust initial pose. Finally, in the fine registration phase, an enhanced Iterative Closest Point (ICP) algorithm is used based on the initial pose. By performing alignment experiments on the Stanford dataset, it is verified that the proposed algorithm has better robustness and can accomplish point cloud alignment efficiently and accurately.