To address the challenge of point cloud registration for complex feature targets， a method leveraging point cloud normal vectors and curvature key features is proposed. Principal component analysis is utilized to compute curvature across varying neighborhood radii， facilitating effective key point selection and initial point cloud downsampling. For each key point， a seven-dimensional feature descriptor is constructed， comprising four normal vectors and three curvature values， thereby encapsulating both angular relationships among normal vectors and curvature characteristics. Similarity between key point descriptors of source and target point clouds is assessed， and correspondences are initially established based on the ratio of the Euclidean minimum distance to the sub-minimum distance. The Random Sample Consensus （RANSAC） algorithm is subsequently employed to eliminate incorrect correspondences and reduce mismatches. High-precision registration is achieved via the Iterative Closest Point （ICP） algorithm， enabling computation of the transformation matrix and quantitative evaluation of registration error. Experimental results demonstrate a root mean square error （RMSE） of 3.32 mm in feature extraction and registration for complex targets， with an average error increment of 0.33 mm/（° ） within a 0-50° registration range. Comparative experiments confirm the superior robustness of the proposed method in large-angle registration of complex targets. Specifically， for space satellite targets， the RMSE of feature extraction and registration is 2.71 mm， accompanied by a Y-direction attitude angle error of 0.427°. The proposed method effectively supports pose estimation and registration of space targets， indicating broad potential for practical applications.