A token-based method for recursively estimating optical flow from a noisy image sequence is presented. Motion characteristics of the perspective projection of 3-D moving object are analyzed and a dynamic model is relevantly built. By taking a pair of or more matching points as the measurement

the recursive estimates of the relevant motion parameters are achieved. Two schemes for motion propagation of the match points

the direct one and the programming one

are discussed

The experiments show that this method needs longer image sequence but rather few match points. It is efficient and robust even in very low SNR conditions.