Multiresolution motion analysis combining feature-based and featureless estimates

In this paper the combination of featureless and feature-based motion analysis is proposed as a way to overcome the limitations that both techniques exhibit when used separately. A multiresolution region-based motion estimator is used to supply affine motion estimates for variable size image regions. These estimates assist the tracking of extracted linear features in the sequence which yield accurate rotational and translational motion estimates for region boundaries. The Multiresolution Fourier Transform is used as a suitable underlying image representation for both parts of the analysis. The feature extraction and tracking process is described in some detail and results are presented, showing the accuracy of the combined motion estimates.