Robustness and accuracy of the existing optical flow algorithms can still be improved for noisy, low textured and poorly contrasted images. Most methods are based on the variational approach given by the Horn-Schunck model which can suffer from over-smoothing problems. In this paper, we propose to take into account the significant structural information of the images during optical flow computation. This will allow for discontinuities across the structure boundaries. In the data term, our optical flow model uses a classical brightness constancy assumption complemented by a modified Hessian constancy assumption based on shape structure. A constraint is added along with the regularization term, for handling the over-smoothing problems. Results on the Middlebury benchmark show the competitiveness of this algorithm. Quantitative optical flow results are given for different noise levels. Finally, application to registration of a complicated bladder cystoscopic data gives a qualitative analysis for robustness of this method.