Affine description of multiple motions in compressed domain

This paper considers the problem of finding a robust solution to the parametric description of multiple motions in video sequences problem, using compressed domain data. Here, we present an effective way to recover an affine global motion description from the MPEG motion information using a robust parameter estimator. The error structure shown by the outliers of this first estimation allows us to cluster them into several local motions, using a non parametric estimator of the density (the mean shift) over a measure of the residuals of our estimation. Finally, results of our algorithm are shown in synthetic and real sequences.