Abstract Aiming at the shortcomings of traditional feature-based coarse registration with low efficiency and many mismatches, we propose a registration method based on feature space matching. We extract the feature space using a simplified PointNet model. We take optimized point cloud PPF information as input and calculate the Euclidean distance according to the extracted feature space vector to filter out matching points. We eliminate the mismatched points to complete the coarse registration through RANSAC, and use ICP to realize fine registration. The results show that the proposed algorithm combined with ICP greatly improves the registration efficiency compared with FPFH and SHOT algorithm, and RMSE of the registration result is smaller.