In order to solve the problems of low efficiency, slow convergence and low accuracy of traditional registration methods, an improved ICP point cloud registration method considering the key points of multi-geometric feature information classification center was proposed. Firstly, the point clouds are downsampled by the voxel grid, and the ground points are removed to speed up the registration efficiency, and then the covariance is used to solve the eigenvalues for the geometric feature information analysis of the non-ground point clouds, and then the Euclidean distance method is used for clustering. The center point of the cluster was extracted as the key point, and the feature key points were described by the FPFH algorithm, so that the center key points of the feature with the same name were correctly paired, and the initial transformation estimation matrix was obtained. Finally, the bidirectional KD-tree and point-to-area ICP algorithm with improved nearest neighbor distance ratio are used for accurate registration, and the Tukey loss function is introduced to resist outlier noise. Compared with the four methods, the results show that the RMSE of the proposed algorithm is 0.202 6 m, which takes 19.426 seconds, and the registration accuracy and efficiency are higher.