Optical-flow estimation of dense motion field using robust techniques

In recent years, interest in motion analysis has increased with advances in processing capabilities. The usual input in a motion analysis system is an image sequence, with a corresponding increase in the amount of processed data. A typical motion problem is to analyze the motion within 2D image data corresponding to a sequence of frames, of a 3D scene. In computer vision a number of techniques are available to estimate the optical flow; the more efficient (in terms of quality) are modulated as the minimization of a global objective function. This cost function includes an observation constraint and a smoothness term. Such models generally assume that the luminance is constant along its trajectory. This assumption is not valid in cases of spatial and temporal distortions as in fluid image sequence. As an extension, a new model is described based on the continuity equation of fluid mechanics and a smoothness function considering the divergence (div) and vorticity (curl) of the motion field. The proposed model is embedded in a multiresolution framework and the minimization is conducted with an efficient multigrid technique. In this paper, the performance of the proposed motion estimation technique is analyzed and compared to similar “standard'” methods, using simulated and real satellite data (the last provided by EUMETSAT). Finally, measures as RMSE of the images, number of cuts and number of iterations for the minimization of the energy function are introduced to justify the improvement of the estimation technique. Key-Words: clouds motion, dense estimation techniques, optical flow, div-curl equations