Efficient motion vector outlier removal for global motion estimation

Because motion vector based global motion estimation methods have much lower complexity than pixel based ones, they are widely used in the compressed domain to estimate the camera motion in video sequences. However, the accuracy of these motion vector based methods largely depends on the quality of the input motion vector field. In real applications, many outlier motion vectors are present because of noise or foreground objects. In this paper, a novel tensor voting based motion vector outlier removal method is proposed to improve the quality of the input motion vector field. First, motion vectors are encoded by second order tensors. A 2-D voting process is then used to smooth the motion vector field. Finally, the smoothed motion vector field is compared to the input one to detect outliers. The experimental results on synthetic and real data show the effectiveness of the proposed method.