Optical Flow at Occlusion

Abstract

We implement and quantitatively/qualitatively evaluate two optical flow methods that model occlusion. The Yuan et al. method [1] improves on the Horn and Schunck optical flow method at occlusion boundaries by using a dynamic coefficient (the Lagrange multiplier α) at each pixel that weighs the smoothness constraint relative to the optical flow constraint, by adopting a modified scheme to calculate average velocities and by using a “compensating” iterative algorithm to achieve higher computational efficiency. The Niu et al. method [2] is based on a modified version of the Lucas and Kanade optical flow method, that selects local intensity neighbourhoods, spatially and temporally, based on pixels that are on different sides of an occlusion boundary and then corrects any erroneous flow computed at occlusion boundaries. We present quantitative results for sinusoidal sequence with a known occlusion boundary. We also present qualitative evaluation of the methods on the Hamburg Taxi sequence and and the Trees sequence.