Accurate Global Motion Estimation Based on Pyramid with Mask

The accuracy of global motion estimation （GME） is usually compromised by the presence of foreground moving objects．This paper presents a GME algorithm with high estimation accuracy．A three-stage mask-based pyramid is built to decrease computational complexity；the initial translation is robustly estimated by a feedback-based algorithm．An approach based on the ratio of outliers in a block and Markov clustering is developed to locate the motion object more effectively than the histogrambased or block-based method；a weighted quadratic error function is adopted instead of the regular truncated quadratic function to increase the robustness to residual error；an improved gradient-based algorithm is introduced to speed up the convergence．From the experimental results on standard video sequences with different motion style, awe conclude that the proposed method could achieve significant performance gains for standard and self-photographed video sequences with camera translation．