Estimation of motion parameters of a rigid body from a monocular image sequence for MPEG-4 applications

We present a method for the estimation of rigid body motion parameters from a monocular image sequence for MPEG-4 applications, such as SNHC face animation. Based on feature extractions in every frame the motion parameters of a human face are estimated with an extended Kalman filter that performs a prediction and correction loop at every timestep. With this recursive structure of the estimation process the temporal redundancies of the motion are taken into account. The nonlinear motion equation is linearized at every timestep within the extended Kalman filter and therefore the rotation is not restricted to be small and the motion model can be based on the first frame and must not describe the frame to frame motion. Results are presented which demonstrate the accuracy of our estimation method on synthetic data as well as on a real image sequence where we estimated the motion parameters of a human face.