Robust multi-frame motion estimation with motion compensated prediction applications

A motion trajectory provides a localised description of motion over multiple frames. This paper describes an approach for estimating a spatially distributed set of such trajectories, with the aim to adequately describe all of the motions present within an image. The problem is conveniently cast in a Markov Random Field (MRF) framework, with a coupled line-process field to handle motion boundaries. Use is made of the resultant set of motion trajectories for the purpose of motion compensated prediction