In Horn and Schunck (HS)'s optical flow method, a weighting parameter α of a motion smoothness constraint plays an important role in determining the estimation accuracy of optical flow. However, conventional methods to optimize the α have been based on an empirical selection or cross-validation, where a global optimization is done at a high computational cost. Thus, a more efficient optimization method is needed. We first assume that real images can be approximated by a two dimensional sinusoidal wave function based on an example of a previously used texture analysis. Two image features of amplitude (the standard deviation of image brightness) and wave number using the sinusoidal wave function are used to analyze and model the relationship between the optimal α and two image features. From the analyzed model, the optimal α can be used to locally minimize the estimation error of optical flow. Because the two simple image features of given real images are used, the optimal α can be efficiently estimated. Experimental results with the estimation accuracy of optical flow show that our proposed method outperforms conventional optical flow ones.
