Bundle adjustment which simultaneously refines motion and structure has the fault of low numerical stability, such as dependence on the initial value, slow convergence speed, and convergence divergence. In this paper, we propose a new multi-frame sequence motion estimation method for stereo visual localization. The method has the fast convergence speed, can converge to the global minimum, and greatly reduce the accumulated error. Stereo visual localization experiments with simulated data and outdoor intelligent vehicle show that our algorithm outperforms bundle adjustment in terms of run-time, accuracy, resistance to noise and dependence on the initial value.