A Method for Extracting and Screening Main Feature Points Based on Salient Geometric Characteristics and NAD Features

Abstract Point cloud alignment is an important task in the field of industrial automation and computer vision recognition. Aiming at the lack of robustness of traditional alignment algorithms in the face of cylindrical objects such as motors or highly symmetric objects, which in turn is prone to poor alignment accuracy or even alignment failure, a method of extracting and screening main feature points based on salient geometric properties is proposed to provide high-precision inputs for point cloud alignment and to improve the position estimation accuracy of symmetric targets. The salient geometric planes and curved surfaces in the target are utilized as the basis of feature point selection to extract more robust main feature points; and different feature descriptors are adopted to describe the feature points based on the target characteristics, which greatly preserves the original main contour and geometric information. A local feature descriptor normalized angle descriptor is designed based on the normal vector, normal angle and Euclidean distance of the point cloud, which is able to effectively remove the incorrect correspondences due to symmetry and feature similarity. Finally, the algorithm for obtaining the global maximum consensus set (GROR) based on the node and edge reliability of the correspondence graph is used to complete the coarse registration, and the iterative closest point algorithm is utilized to complete the fine registration. Experimental results on motor point clouds taken from different viewpoints show that the proposed registration strategy is visually and numerically superior to existing state-of-the-art methods, especially when there is only a single-frame point cloud of the target.