Robust optical flow from photometric invariants

Abstract

Optical flow is widely in use in the field of image processing. In general, optical flow is computed from luminance images. However, optical flow based on luminance information highly depends on moving shadows, varying shading and moving specularities due to camera movement, and fluctuations in the light source intensity. In this paper, we propose a novel method to compute optical flow based on photometric invariants. A major drawback of photometric invariants and their derivatives is that they are unstable for certain RGB values. Therefore, we study on photometric invariant derivatives and noise propagation yielding a condence measure indicating the stability of the corresponding photometric invariant derivatives. This condence measure is integrated into the optical flow framework to provide robustness against noisy data. Experimental results show that the proposed method significantly outperforms optical flow estimation that does not take the instability of the invariants into account.