Previous motion-estimation methods, such as the multi-resolution
Laplacian pyramid technique, are restricted in the magnitude of
recoverable motion by the width of the spatial derivative operator
at the top of the resolution pyramid. Larger motions are often corrupted
by aliasing. Increasing the number of levels of resolution tends to blur
object structure and merge regions undergoing different motions,
reducing the accuracy of motion estimation. In addition, such techniques
treat the computation of motion as a set of independent processing
problems between pairs of consecutive frames.
A spatio-temporal approach to motion estimation has been described here that exploits the gradual variation of visual motions through a sequence. Motion estimates derived for one image pair are fed forward and used as initial estimates for the next. The novel estimator derived in section 2 can embrace more general spatio-temporal models that describe motion over long sequences of frames using single sets of motion parameters. Moreover, the motion parameters produced by this method can recover more complex motions, such as curved trajectories and changes of perspective, which cannot be modelled by the multi-resolution approach or block matching methods.
Graeme Jones