Smoothing of optical flow using robustified diffusion kernels

This paper proposes a new optical flow smoothing methodology combining vector diffusion and robust statistics. Vector smoothing using diffusion preserves moving object boundaries and the main motion discontinuities. According to a study provided in the paper, diffusion does not remove the outliers but spreads them out, introducing a bias in the neighbourhood. In this paper robust statistics operators such as the median and alpha-trimmed mean are considered for robustifying the diffusion kernels. The robust diffusion smoothing process is extended to 3-D lattices as well. The proposed algorithms are applied for smoothing artificially generated vector fields as well as the optical flow estimated from image sequences.