Non-Rigid Point Set Registration via Mixture of Asymmetric Gaussians with Integrated Local Structures

As a fundamental problem in computer vision community, non-rigid point set registration is a challenging topic since the corresponding transformation model is often unknown and difficult to model. In this paper, we present a robust method for non-rigid point set registration. Firstly, a mixture of asymmetric Gaussian model (MoAG) is employed to capture spatially asymmetric distributions which the Gaussian mixture model (GMM) based methods neglect instinctively. Secondly, local structures among adjacent points are integrated into the MoAG-based point set registration framework to improve the correspondence estimation. Thirdly, Expectation-Maximization (EM) algorithm which provides a numerical method for finding maximum likelihood estimators is utilized to estimate parameters of the latent variable model in our proposed method. The transformation model is solved under regularization theory in the Reproducing Kernel Hilbert Space (RKHS). Conveniently, a fast implementation is achieved by the sparse approximation. Finally, experiments results show that the robustness of our algorithm outperforms the comparative state of art methods under various types of distortions, such as deformation, noise, outliers and occlusion