P.R. Giaccone and G.A. Jones
School of Computer Science and Electronic Systems
Kingston upon Thames
Surrey KT1 2EE, United Kingdom.
Spatio-temporal extensions to the classical optical flow technique are presented. A novel optical flow estimator is derived capable of incorporating spatio-temporal motion models by iteratively refining initial motion estimates. These initial estimates may be derived from motion in previous frames using a spatio-temporal processing architecture, or from block matching or feature matching techniques. By incorporating spatio-temporal motion models with explicit temporal components, a single set of motion parameters can describe potentially curved pixel trajectories over a sequence of frames. In addition, given approximate initial estimates, the method can handle arbitrarily large visual velocities without recourse to multi-resolution techniques. Results are given showing that the latter method can recover complex motions that purely spatial methods cannot. Finally, an outline is given of applications of the spatio-temporal motion model in the fields of cinematography and video compression.