Kalman filter formulation of the nonuniform image motion estimation in transformed-domains

This paper is concerned with the development and implementation of a new pixel-recursive algorithm for nonuniform image motion estimation. Using a unitary transformation, the motion vector is described by a coefficient vector. Since the motion field is typically low-pass, the dimension of the coefficient vector is significantly reduced; the resultant truncated and constant vector is then easily modeled in state-space. A Kalman filter is subsequently implemented to estimate the motion coefficients. Additionally, the paper presents a detection scheme for identification of (detectable) moving pixels from noisy data using a composite likelihood ratio test (LRT). Simulation experiments are presented using real images; results of the simulations are presented