Non-Rigid point set registration is very important for many fields of study. Currently, the famous algorithms generally use correspondence estimation and transformation update based on single feature and single constraint. But performance and application area of the single feature and constraint based algorithms are limited. This paper presents a non-rigid point set registration method based on dual-feature Gaussian mixture model and dual-constraint transformation. Firstly, a dual-feature descriptor is defined and global feature and local feature are used to build the dual-feature descriptor. Then, Gaussian mixture model is improved to obtain a dual-feature Gaussian mixture model by the dual-feature descriptor. Finally, a local structure constraint descriptor is defined and used together with global structure constraint descriptor to preserve the local and global structures of point set. A method is presented for running estimate correspondence that uses dual-feature Gaussian mixture model and updates dual-constraint transformation based on Gaussian radial basis function iteratively to match non-rigid point set accurately. Performance of the presented method is evaluated by synthetic point set registration, CMU sequence image registration, remote sensing registration, IMM face data registration and true image feature point registration. Comparing with other eight state-of-the-ate methods, the new method shows the best alignments in most scenarios.