Research On Point Set Registration Algorithm Based On Structural Constraints And Gaussian Mixture Model

As a key topic in the field of computer vision,point set registration has been widely used in numerous occasions,such as defect detection,medical image analysis,and so on.The goal of point set registration is to find an optimal spatial transformation to align a template point set to a target point.Due to some unfavorable factors,such as unknown non-rigid spatial transformations,noise,and outliers,point set registration has become a challenging problem.To address these issues,the following two works have been carried out in this thesis.(1)We propose a robust point set registration method based on local deviation constraint and smooth constraint.In many existing algorithms,the estimation results of some points may deviate greatly from their true points,because only the overall registration error is considered in their models.To address this issue,a local deviation constraint term is designed,by using the deviation of registration error between the transformed model points and the target points.In addition,in order to smooth the process of spatial transformation,the square norm of the reproducing kernel Hilbert space is used as the regularization term of the Gaussian mixture model.Finally,the correspondence matrix is computed by a binary representation via the minimum spanning tree induced triangulation algorithm.Compared to a probability matrix,the corresponding matrix obtained by this method has a better estimation performance.The experimental results on several widely used data sets demonstrate the effectiveness of the proposed method.(2)An effective representation of membership probability is proposed for each component of the Gaussian mixture model.The proposed representation can better capture the probability difference between the matched points and the unmatched points.As a result,the probability representation becomes more accurate for the Gaussian mixture model.To make full use of the structure information of points,a new membership probability representation for Gaussian mixture model is proposed,by utilizing the feature descriptor of shape context or fast point feature histograms.Moreover,considering each point of the model points,a dynamic programming algorithm is used to find the optimal candidate points from the target points.Experimental results on several widely used 2D and 3D data sets show that the proposed approach is more robust to deformation,outlier,occlusion,and rotation.