A Robust Optical Flow Computation under Variational Illumination Based on Improved Census Transformation

In wild environment, the optical flow algorithm accuracy will be affected by the changing illumination. In order to solve this problem, a variational optical flow computation algorithm based on improved Census transform is proposed in this paper. In traditional Census transformation, different sub-images cannot be distinguished due to insufficient amount of information. In this algorithm, completed gradient flow descriptor in an image window is constructed to overcome this problem. Furthermore, transformation window is selected adaptively according to image contour which can improve the description accuracy in depth discontinuity areas. For building a binary string, the elements which are closer to the reference element are set in the higher order of binary string, weakening the impact of image distortion to the optical flow estimation. The improved describer is used to construct the data term in TV_L1 variable flow computation framework. Gaussian pyramid decomposition and median filtering methods are also used to estimate optical flow in the solving processing. The proposed improved Census transformation contains more local image characteristics, which has the stronger ability to resist illumination change. The proposedalgorithm is tested on the benchmark of both Middlebury and KITTI comparing to the traditional Census transformation method. The experimental results demonstrate the effectiveness of the proposed method.