Motion field modeling for motion estimation and compensation

This paper proposes a new way of representing motion fields and develops a predictive model for the motion field of video sequences exploiting this representation. Generalizing linear prediction models of time-series, the model characterizes the relative importance of neighboring pixels in estimating the motion at a given pixel. The model is applied to develop a weighted pel-recursive algorithm for motion estimation, improving the performance of basic pel-recursion. An extension of the basic model, exploiting the time-reversibility of motion, produces further improvements