It is currently recognized that optical flow computation has many applications in image processing, pattern recognition, data compression, and biomedical technology. Differential approaches, which estimate velocity vectors from the spatial and temporal intensity derivatives are the most common visited techniques in the literature. The basic assumption behind that is the intensity variations are weak and only due to a movement in the image plan.
This constant brightness assumption leads to an ill-posed problem that can only be solved by imposing additional assumptions. A standard technique, which is due to to Horn and Schunck, is to require the flow field to be smooth by means of a standard regularization approach. This results in a system of elliptic PDEs of reaction diffusion type. The second order terms are induced by the regularization and become straightforward Laplace operators. The zero order terms are linear (but variable) and are computed from the derivatives of the image data. Since the image data (and even more so its derivatives) are usually nonsmooth, this poses some nontrivial problems. The PDEs are usually discretized by standard finite differences.
The Horn-Schuck algorithm is the standard way to discretize and solve the PDEs. In multigrid terminology, it is simply a coupled pointwise relaxation. As expected, it has acceptable convergence only if the zero-order terms dominate but the performance is poor, when the diffusion character dominates. In this case, a multigrid strategy can be expected to provide a significant accelleration by allowing a quick propagation of information from non-zero flow field regions into homogeneous or untextured image areas. The Horn-Schunck algorithm can still be used as a smoother for such a multigrid method. The most difficult aspect is to deal with the strongly discontinuous coefficients in an efficient way. This will be discussed in our presentation.