Motion Estimation Based on Patchwise-Optimized Optical Flow

Abstract

Optical flow has been a promising technique for motion estimation from video. Existing optical flow methods primarily focus on three constraint terms between two frame images, which are brightness constancy, gradient constancy and smoothness constancy for each pair of corresponding pixels. However, such pixelwise constraints are not enough to summarize the local motions for most of objects and thus usually cannot achieve high accuracy in the motion estimation. In this paper, we present a novel framework to optimize the optical flow using patchwise brightness constraint, gradient constraint, as well as smoothness constraint. Since patchwise constraints can describe more local statistics invariant motion features for the objects located within the windows, we can obtain an optical flow technique with more stable and accurate performances. Our method is evaluated on a standard dataset. Experimental results show that our patchwise-constrained optical flow generally outperforms existing pixelwise-constrained optical flow.