Point clouds registration with surfaces of low curvatures

In this paper, we propose a novel method for point clouds registration. In point clouds, there always exist a large number of surfaces with low curvatures. Let the surface locate in an OXYZ coordinate system, if we rotate the mean normal of the surface parallel with the Z axis, the 3D surface can be projected to 2D image via orthographic projection. Then we can detect and describe SIFT features in the 2D image. For two point clouds, we can find matched features between these 2D images. As we know the relationship between 2D images and 3D point clouds, the corresponding 3D points between different point clouds can be obtained. Further, we use the SVD method to estimate the transformation parameters between two point clouds. Moreover, in order to obtain the parameters robustly, the RANSAC method is combined to reject error matchings. The proposed method is able to register the point clouds effectively, and it is demonstrated in experiments.