Robust computation of optical flow field with large motion

The robust computation of large motion has been an important but difficult problem in dynamic scene analysis. This paper presents a novel algorithm to determine optical flow field with large motion from the sequences of images. The algorithm integrates the modified gradient constraint equation with the anisotropic diffusion, which not only decreases the error of constant image brightness assumption caused by large motion, but also alleviates the error propagation effect. The algorithm is based on a multi-scale Gaussian pyramid. The large-scale intensity information at coarse image level is first used to obtain rough estimations of image motion, which are then refined by using the small-scale intensity information at fine image level. Experiments on synthetic and real image sequences have shown that the method had good computation accuracy and robustness for large motion.