Flexible global motion estimation oriented to video object segmentation

This paper presents a flexible global motion estimation (GME) algorithm based on gradient methods, which can be used for video applications, such as sprite coding. This approach improves the precision of existing GME algorithms by introducing robust initial translation estimation and adaptive technique to remove the influence of outliers. The method excludes the outliers using flexible threshold and a weighted quadratic error function is adopted instead of regular truncated quadratic function. Three-stages pyramid is built to increase robustness and fast estimation algorithm is introduced to decrease computational complexity. From the experimental results we conclude that the proposed method provides significant performance gains for video sequences with camera translation.