Model based motion estimation using constrained weighted least squares

Motion estimation from video sequences is important in many applications such as image coding, visual navigation, and security monitoring. The apparent deficit of existing methods have been their sensitivity to the presence of outliers and their limited ability to deal with multiple moving objects. In this paper, we present a motion estimator which models local image flow field using second-order polynomials with 6 parameters. We also show how constraints on the magnitude of image velocities as well as an the projective transformation can be used to better compute image flows even at the presence of outliers as well as the existence of multiple moving objects. Image flow results for several red video sequences show the robustness of this estimation scheme