Model-Based Motion Estimation and Its Application to Restoration and Interpolation of Motion Pictures.
Abstract:
This report is concerned with the problem of motion estimation from digitally sampled motion pictures. Several models are developed that describe object motion with velocity fields. Given an image sequence, the velocity field is underconstrained and therefore cannot be determined uniquely. However, by imposing structural constraints on the velocity field in the form of a parametric model, it is possible to determine the model parameters uniquely. This report describes parametric models which form the basis for two motion estimation algorithms. Experimental results demonstrate that these algorithms determine velocity fields more accurately than conventional region matching methods. One of the algorithms, based on the least squares error criterion, also has the desirable property of being computationally efficient. To demonstrate the performance of the least square motion estimation algorithm, a motion-compensated noise reduction system was implemented. Experiments demonstrate that the motion-compensated noise reduction system can yield better results than conventional restoration methods. A motion-compensated frame interpolation system, also implemented, permits frame rate conversion by arbitrary rates.