The motion estimation in stereo visual localization is usually carried out in 3D Euclidean space. 3D points are corrupted with heteroscedastic and anisotropic noise, and the depth direction of 3D reconstruction is less accurate than the two other direction. And therefore, the accuracy of motion estimation in 3D Euclidean space is low. We propose a new high-precision stereo visual localization algorithm based on motion estimation in disparity space. Initial robust motion parameters and matching inliers are obtained by four points closed-form linear motion estimation in disparity space combined with RANSAC algorithm. Then, we propose nonlinear motion optimization method based on L-M algorithm in the new disparity space re-projection error function. The mothod is appied to optimise the initial motion parameters. The noise in disparity space is homogeneous and isotropic, and our initial linear motion estimation and nonlinear optimization method all are carried out in disparity space and can achive globally minimum. Simulated and real experimental result show that our method can achive higher localization accuracy than traditional motion estimation method in 3D Euclidean space.