Using Fourier local magnitude in adaptive smoothness constraints in motion estimation

Like many problems in image analysis, motion estimation is an ill-posed one, since the available data do not always sufficiently constrain the solution. It is therefore necessary to regularize the solution by imposing a smoothness constraint. One of the main difficulties while estimating motion is to preserve the discontinuities of the motion field. In this paper, we address this problem by integrating the motion magnitude information obtained by the Fourier analysis into the smoothness constraint, resulting in an adaptive smoothness. We describe how to achieve this with two different motion estimation approaches: the Horn and Schunck method and the Markov Random Field (MRF) modeling. The two smoothness constraints obtained are compared with standard solutions. Experimental results with synthetic and real-life image sequences show a significant improvement of motion estimation in both cases.