Motion estimation and tracking using multiresolution affine models

As in many areas of image analysis, a key issue when tackling motion is determining the appropriate size of local analysis window to be used. The work described in the paper attempts to address this problem. It is based on two key elements: the use of an affine model to describe local motion variation within a sequence; and the incorporation of this local model into a multiresolution framework to describe the global motion field. Affine models provide greater flexibility in modelling local motion, being able to represent rotation, dilation and shear as well as translation, enabling larger window sizes to be used and hence the potential for more robust estimation. The use of a multiresolution framework allows the window size to be adapted to the underlying motions present in the scene, providing a mechanism for selecting the best set of local affine motion descriptors. The result is a piecewise-linear representation of the motion field in terms of a 'mosaic' of 'patches' of varying size moving with affine motion, corresponding to planar facets moving with rigid motion in 3-D, the so-called 2.5-D model. The approach has led to the development of algorithms for estimating motion between frames and tracking region elements over time. The algorithms are efficient and lead directly to concise descriptions of the motion field in terms of a relatively small number of mosaic elements. Experiments have shown that the estimator and tracking mechanism perform well on both synthetic and natural sequences.