Optical flow techniques are used to compute an approximate motion field in an image sequence. We apply a variational model for the optical flow introducing a curvature based regularizer that requires the solution of a fourth order system of partial differential equations with jumping coeffcients. A geometric multigrid solver for that problem is presented which is composed of collective Gauss-Seidel relaxation and standard geometric transfer operators. Galerkin based coarse grid operators are applied for an effcient treatment of jumping coeffcients. Finally, some results on convergence rates, timings and visual quality of the approximated motion field for synthetic and real world images are shown.