A New M-Estimator Approach for Global Motion Estimation

Global motion estimation (GME) is an extensively used tool in a variety of video processing applications. The major challenge in this field is the presence of large foreground objects. There is a wide variety of algorithms addressing this problem. The major shortcoming of these algorithms is inconsistent performance over several video sequences. In this paper, we propose a GME approach that is fully automatic, can successfully handle large foreground objects and provides consistent results over a range of different video sequences. The proposed method initially coarsely determines the foreground pixels by a clustering technique. The effect of remaining foreground pixels in the estimation process are then reduced by using a modified Lorentzian estimator. Experimental results prove the superiority and consistency of the proposed method compared to some recent approaches.