Non-rigid point set registration based on DIS&ANG descriptor and RANSAC

Point set registration is a general problem in many domains, such as 3D reconstruction, simulation design, computer vision, and motion tracking. For non-rigid transformation, the local structure of the points reserve relatively complete, and has important significance. But the general point set registration algorithm only considers the global structure information, the local structure information is given the air. In this paper, we propose a local structure descriptor, which not only considers the distance information of the local neighborhood, but also considers the angle information between neighbors. Therefore, the description operator can more accurately describe the local structure information. Combined with the above descriptor, a non-rigid point cloud registration algorithm based on the random sample consensus (RANSAC) and the local structure information is proposed in this paper. Firstly, we use the random sample consensus algorithm to make an initial registration. Secondly, the local structure descriptor and the Gaussian mixture model are used to register exactly. Extensive experiments show that our algorithm has obvious improvement than the state-of-the-art methods under various types of distortions, such as deformation, outliers and rotation.