Motion estimation using adaptive correlation and local directional smoothing

We present two forms of correlation for addressing the problems presented by non-translational and multiple motions in region based motion estimation. The first-registered correlation-enables affine motion parameters to be estimated; whilst the second-iterated partial correlation-enables the identification and estimation of two motion components within a region. Incorporation of both techniques in a multiresolution framework provides for more flexibility in region selection in terms of size versus robustness. We also present a local directional smoothing algorithm based on certainty measures obtained from the correlations which further improves the motion estimation. Results of experiments involving both types of region motions illustrate the effectiveness of the three schemes.