On improving the robustness of variational optical flow against illumination changes

The brightness constancy assumption is the base of estimating the flow fields in most differential optical flow approaches. However, the brightness constancy constraint easily violates with any variation in the lighting conditions in the scene. Thus, this work proposes a robust data term against illumination changes based on a rich descriptor. This descriptor extracts the textures features for each image in the two consecutive images using local edge responses. In addition, a weighted non-local term depending on the intensity similarity, the spatial distance and the occlusion state of pixels is integrated within the adapted duality total variational optical flow algorithm in order to obtain accurate flow fields. The proposed model yields state-of-the-art results on the the KITTI optical flow database and benchmark.