Michael Bleyer, Margrit Gelautz, and Christoph Rhemann
This paper describes an algorithm for computing the optical flow field between two consecutive frames. The algorithm takes advantage of image segmentation to overcome inherent problems of conventional optical flow algorithms, which are the handling of untextured regions and the estimation of correct flow vectors near motion discontinuities. Each segment's motion is described by the affine motion model. Initial motion segments are clustered to derive a set of robust layers. The assignment of segments to layers is then improved by optimization of a global cost function that measures the quality of a solution via image warping. Occlusions in both views are detected and handled in the warping procedure. Furthermore, the cost function aims at generating smooth optical flow fields. Since finding the assignment of minimum costs is NP-complete, an efficient greedy algorithm searches a local optimum. Good quality results are achieved at moderate computational expenses.
In Joint Hungarian-Austrian Conference on Image Processing and Pattern Recognition