A multiscale technique for optical flow computation using piecewise affine approximation

We present a technique to estimate the optical flow in an image sequence, based on a piecewise affine model. In this piecewise approach, the area of interest in each image frame is divided into a set of small triangular patches. These triangular meshes are established over a set of feature points, which are extracted from the images and tracked from one frame to another. The velocity field within each triangular patch is parameterized by an affine transform. A multiscale coarse-to-fine approach is employed to increase the robustness of the method as well as the accuracy of the optical flow resulted from piecewise affine approximations. Finally, an adaptive filter is used to refine the estimated flow field. The filter is designed in such a way that not only can it reduce noises caused by errors of the process described above, but it can also avoid smoothing the discontinuities in the motion field. The method has been implemented and some experimental results are presented in this paper. The method takes advantage of widely used MPEG-4 encoding hardware/software tools