Abstract: The ICP registration algorithm has high requirements for the initial position of point clouds and low registration ability for point clouds with low overlapping rate.To solve these problems, a point cloud registration method based on feature matching of key points was proposed.A key point extraction algorithm based on the differences of mean values of multiscale weighted normal projection was designed, and then key points were characterized by SHOT descriptors.Fusing of geometric consistency and RANSAC algorithm, the mismatched point pairs were removed in the matching process, and the correspondences between key points were optimized.The transformation matrix was obtained by using singular value decomposition, cloud rough registration was completed, and fine registration was performed by using ICP.The experiment results show that the key point extraction algorithm proposed in this paper can effectively extract points with obvious changes in the surface of the point cloud.And using SHOT descriptor to characterize the key points can complete registration of point cloud data quickly and accurately.At the same time, for the point cloud with low overlapping rate, the proposed method also has better registration effect.