In view of the problem of low accuracy and poor stability of 3D space registration in mobile augmented reality, an improved registration method based on LMeds (Least Median Squares) is proposed in this paper. The method first uses the ORB (Oriented FAST and Rotated BRIEF) algorithm to carry out the feature detection, and the FREAK (Fast Retina Key-point) algorithm is employed to describe the features. Then the feature points are screened by the hamming distance. In order to obtain good matching points, the Sampson distance estimation is introduced in the LMeds algorithm. By eliminating the mismatch points in time to reduce the cumulative number of calculation cost functions, it not only maintains the good robustness of the LMeds algorithm, but also effectively reduces the computational complexity of the algorithm. Finally, camera pose estimation is conducted to get the absolute attitude of video image relative to the identified image, so as to enhance the scene. Experimental results show that this method improves the precision of feature matching and the accuracy of 3D registration, and also has good real-time performance and robustness when environment changes.