Window design for overlapped block motion compensation through statistical motion modeling

This paper presents an analysis of the block-decimated motion estimates and relates them to the underlying motion random field. It further parameterizes the scene intensity random field and the motion random field in terms of their correlation properties. Within this framework, we develop an algorithm to optimize the window for overlapped block motion compensation as a function of the model parameters. Through simulations, we demonstrate that the optimal window resulting from the parametric formulation offers performance comparable to the window deterministically optimized for the test sequence, and it offers more robust performance outside the training set. Finally, we apply our algorithm to adapt the overlapped window to match the temporally changing characteristics of the scene and motion fields. We demonstrate that for real-time applications, where the number of frames used for adapting the window is limited, our algorithm significantly outperforms the method introduced by Orchard and Sullivan (see IEEE Trans. Image Processing, vol.3, no.5, p.693-9, 1994).