A new algorithm for estimating optic flow for low-level vision systems

A method of estimating the optical flow in a noisy image sequence is presented. In this approach a gradient matching procedure is effectively combined with an image differencing technique to obtain displacement vectors between two successive frames. The resulting displacement field is dense, and leads to compact depth maps. The computational cost is shown to be lower than in a simple correlation method. A maximum likelihood estimate is used to characterise the probability of match which can give a local estimate of the noise distribution in the frame sequence. Such knowledge can be used effectively for the determination of displacement vectors in subsequent frames of the image sequence.