Global Motion Estimation Based on the Multi-stage Analysis of Motion Vectors

Global motion estimation based on motion vector field has lower complexity than pixel-based method, so it is widely used in video segmentation and compression. However, outlier motion vectors, caused by image noise or foreground objects, reduce the accuracy of motion vector-based global motion estimation. In this paper, a global motion estimation algorithm based on the motion vector multi-stage processing is proposed to improve the estimation accuracy. The proposed method adaptively removes foreground objects by comparing the motion characteristics differences between the local motion and global motion area. For each block considered, the motion similarity between the neighboring blocks is exploited to detect the cycle smooth area. The isolated noise area is also filtered out. Finally, the inlier motion vectors are used to estimate the global motion parameters. Experimental results show that the proposed scheme filters effectively outlier motion vectors and improves the accuracy of global motion estimation.