Global motion estimation (GME) is widely used in image/video processing and various applications. But the accuracy of
estimation results is badly influenced by local motion and noises. Furthermore, the conventional GME algorithms in
spatial domain usually need a large number of iteration times, which makes computational complexity extremely higher.
In this paper, we propose an efficient and fast GME algorithm based on motion vector field, which adaptively selects
input pixels for solving transform models. More characteristics of the image are considered, such as the difference
between global motion and local motion, the distribution of motion vectors, and macroblock partition modes. The
proposed algorithm includes three steps: First, we obtain several sets of pixels by merging similar bins in the histogram
of motion vectors and generate a weight map. Second, we choose the cluster with the minimum distribution variance in
the image as the cluster representing the global motion. The pixels with higher weights in this cluster are chosen as the
input pixels for solving transform models. Finally, we employ the 6-parameter affine model as the transform model and
calculate the parameters. Experimental results show that the proposed algorithm is effective and fast.