Occlusion-Aware TV-L۱ Optical Flow with Second-Order Regularization

We introduce a new variational framework for optical flow estimation that jointly accounts for occlusions and smooth yet discontinuity-preserving motion fields. Classic TV-L۱ approaches often struggle where brightness constancy fails—such as at occlusions—and may produce undesirable “staircase” artifacts due to first-order total variation regularization. To tackle these limitations, we propose an occlusion-aware formulation in which each pixel is assigned a mask variable indicating whether it should adhere to brightness constancy. Simultaneously, we replace the standard TV term with a second-order total generalized variation (TGV) prior, allowing the estimated flow to transition smoothly within objects while preserving sharp boundaries. The resulting convex energy is efficiently minimized via a primal-dual algorithm, embedded in a coarse-to-fine warping strategy for handling large displacements. Experiments on synthetic and real datasets, including Middlebury and KITTI, show that our method outperforms classical TV-L۱ in occluded areas and yields more realistic flow fields with reduced staircasing. These results demonstrate the benefits of pairing explicit occlusion handling with second-order regularization in a unified optical flow framework.

کلیدواژه ها:

Optical Flow Estimation ، Occlusion Handling ، TGV ، TV-L۱ Regularization ، Motion Discontinuities