Estimating optical flow for large interframe displacements

Abstract

We present a new algorithm for estimating optical flow in the presense of large interframe displacements. This algorithm relies on the clustering of velocities of spatioternporally neighboring object points in velocity-time. Our algorithm is able to a large extent to overcome the aperture problem [5] and the problem of occlusion [5], Our algorithm has advantages over prior algorithms in situations where a) the cinematic sequences being dealt with have been sparsely sampled spatially and/or temporally b) there are large interframe displacements — of the order of several pixels, c) the intensity distribution in the images is non-linear, and d) the optical flow field exhibits discontinuities which need to be accurately detected. We present the results of some experiments conducted to test the performance of our algorithm on real cinematic data, and we compare this performance to that of two other algorithms on the same data.