Multiresolution Motion Estimation Using an Affine Model

This report describes a novel method for estimating motion in image sequences. The approach is based on a multiresolution model in which the motion eld is considered to consist of a collage of di erent size regions with di erent local motions associated with them. The local motions are de ned in terms of an a ne model. The estimation procedure consists of a local component in which the a ne motion parameters are estimated using a generalised correlation technique and a global component in which the nal motion eld is determined using a coarse-tone matching scheme. E cient implementation of the scheme is achieved by the use of a generalised wavelet transform, the Multiresolution Fourier Transform (MFT). Results of experiments performed on both arti cial and natural image sequences illustrate the potential of the approach.