Abstract
One of the most important step in precisely optical flow computation is using L1 norm for mathematical modeling. For discrete signals, L1 norm gives better results than L2. Another useful ingredient is using local and global combinations in order to use the advantages of both methods. We will present a combined local-global approach that uses L1 norm for both the data fidelity terms and the regularization one. Our approach is robust to noise and occlusions and preserves motion boundaries. Additionally, a version using L1 only for the regularization term and another one using only L2 are presented. We will show that combined local-global estimators have some benefits in real scenarios. All numerical schemes resulted are highly parallelizable, being designed for running on graphic processing units.