Differential optical flow techniques estimate flow fields based on the derivatives of consecutive images. However, the use of partial derivatives amplifies the possible noise present in those images, thus degrading the accuracy of the computed flow fields. This problem is usually overcome by smoothing the gradient images with Gaussian filters. However, the latter tends to blur discontinuities, yielding an undesired loss of accuracy. This paper proposes tensor voting as an alternative to Gaussian filtering that yields more robust and accurate optical flow fields. The proposed model yields state-of-the-art results on the Middlebury optical flow database and benchmark.