An optical flow method combining Gaussian convoluted data term with non-local median filter is proposed to remove noise and consequently improve the robustness and accuracy of the optical flow estimation. Robust L1 norm is employed for construction of data term, which is smoothed with Gaussian filter to suppress noise, and primal-dual method is introduced to improve the estimation efficiency of variational optical flow. A global optimization strategy based on non-local median filter is used to further enhance the estimation accuracy. The coarse-to-fine pyramid technique is employed to improve the adaptability of the algorithm for large displacements estimation. The proposed method is evaluated by using both the Middlebury optical flow database images and real scene images. The experimental results show that the proposed method performs good robustness and accuracy in contrast with traditional TV-L1 model algorithms.