Optical flow estimation via weighted guided filtering with non-local steering kernel

摘要

Abstract The weighted median filter and the guided image filter are considered important methods for the recently popular variational and non-local total variational optical flow estimation. Their attractive advantages are that outlier reduction is attained, while motion boundaries are preserved. However, these methods still suffer from halo artifacts near edges caused by motion occlusion and illumination changes in adverse outdoor conditions. To overcome these drawbacks, we propose weighted guided filtering with a non-local steering kernel during the coarse-to-fine optical flow estimation. The weighted guided filtering can preserve the motion edges more efficiently by incorporating edge-aware weighting into the filtering process, and the non-local steering kernel can leverage the edge direction more sufficiently. First, we formulate weighted guided filtering with a non-local steering kernel to preserve the edges and improve the robustness of optical flow estimation. Second, we present a combination of median filtering and weighted guided filtering with a non-local steering kernel to optimize the optical flow estimation under the coarse-to-fine process. We compare the proposed method with several state-of-the-art methods using the Middlebury and MPI Sintel test datasets. The results indicate that the proposed method is robust for optical flow estimation and able to preserve motion boundaries.