Global motion estimation using modified least median of square error

Global motion estimation is an important step in video coding. It is needed in automatic video object segmentation for content-based video coding and useful in motion compensation. Among the many methods for global motion estimation, the least median of square error (LMedS) is one of the best because of its robustness in outlier rejection. However, its performance deteriorates if the number of error medians computed is not sufficient or outliers occupy more than 50% of the total pixels. In this paper, we show how the number of medians affects the performance and that allows one to obtain a good tradeoff between performance and computation requirement. Also, we have developed a new method that allows LMedS to maintain its excellent performance when outliers occupy more than 50% of the total pixels.