Optical Flow Estimation from Layered Nearest Neighbor Flow Fields

Abstract

The goal of this paper is to study a strategy of preserving local weak motion regions in large displacement motions variational optical flow methods. We review the development of optical flow estimation for large displacement motions, and then point out to the limitation of the existing method in the local weak motion regions. Finally, we present a layered nearest neighbor flow fields estimation method to overcome this drawback. we perform pyramid layering on the input image sequences and use TreeCANN method to obtain the nearest neighbor field between each layer of images. Then, according to the dominant motion modes in the nearest neighbor field to performe motion segmentation refinement and to get the nearest neighbor flow field. Finally, the nearest neighbor flow field is integrated into the non-local total variational with $L^{1}$ norm (TV- $L^{1}$ ) flow fields estimation model to recover the flow field detail information of the local weak motion regions lost in the TV- $L^{1}$ model. The test results prove that our method can effectively recover the flow field in the local weak motion region and has good accuracy and robustness.