A novel motion vector outlier removal technique based on adaptive weighted vector median filtering for global motion estimation

In this paper we propose a novel motion vector outlier removal technique for global (camera) motion estimation based on adaptively weighted vector median filtering. The accuracy of motion vector based global motion estimation algorithms is highly dependent on the ability of the system to reject outlier motion vectors. The outlier motion vectors may be due to noise, foreground objects or due to the encoders compression requirements. Our idea is based on the premise that by minimizing the effect of outlier motion vectors, the efficiency of the global motion estimation algorithms can be improved. In our work, the adaptively weighted vector median filter is used to smoothen the motion vector field followed by comparison of the smoothed motion vector field with the input motion vector field to detect the outliers. The detected outliers are then excluded from the global motion estimation process to get a robust estimate of the camera motion parameters. We compare our proposed method with existing outlier rejection techniques using both synthetic as well as real video sequences to show the effectiveness of our proposed method.