Refinement of optical flow estimation and detection of motion edges

In this paper a multi-scale method for the estimation of optical flow and a simple technique for the extraction of motion edges from an image sequence are presented. The proposed method is based on a differential neighborhood-sampling technique combined with a multi-scale approach and flow filtering techniques. The multi-scale approach is introduced to overcome the aliasing problem in the computation of spatial and temporal derivatives. The flow filtering is useful near motion boundaries to preserve discontinuities. A residual function, which is a confidence measure of the least-squares fit used to compute the optical flow, is introduced and used to filter the flow and to detect motion boundaries. These boundaries, that we call motion edges are extracted by searching for the directional maxima of the map obtained by thinning this residual function. The proposed method has been tested in a variety of conditions. The results obtained with test images show that the proposed approach is an improvement of previous techniques available in the literature.