Motion estimation algorithm of stereo vision is one of the important factors which affect the accuracy of stereo visual localization. The traditional 3D-3D motion estimation is greatly affected by noise, so the accuracy of algorithm is not high. We propose a new stereo visual localization algorithm based on 2D-3D binocular motion estimation. In our method, 2D image projection coordinates is directly used instead of 3D coordinates of feature after motion. Firstly, EPnP motion estimation is applied to determine matching inliers and initial motion parameters. We propose 2×2D-3D motion parameters optimization method based on Levenberg-Marquardt algorithm and geometric constraints between the 2D projection of binocular cameras. Optimal 2D-3D motion estimation is achieved by minimizing the error between the observed 2×2D image points and the re-projected 2×2D image points of stereo cameras from the reconstructed 3D features. Simulated experiment and outdoor real experiment show that our method performs well in accuracy and robustness, and are better than traditional 3D-3D methods.