On illumination-invariant variational optical flow for weakly textured scenes

Abstract

This paper deals with variational optical flow approaches for motion estimation under varying illumination conditions in weakly textured scenes. It proposes a systematic and complete study on descriptor-based data-terms that lead to a robust variational optical flow model. Unlike the literature which most often only experimentally shows that a descriptor is illumination invariant, this contribution gives a theoretical proof of this invariance. First, a local illumination change model is proposed and used to mathematically check whether a descriptor is invariant or not with respect to illumination variations between images. Then, this contribution proposes two general mathematical formulations which can be used to design a wide variety of new illumination-invariant descriptors. To illustrate the interest of the proposed approach, two novel illumination-invariant descriptors are constructed using the proposed general formulations. Moreover, the performance of the descriptors was evaluated on numerous datasets with known ground truth optical flow, while the robustness of the variational optical flow approach was highlighted using complex medical image sequences without ground truth. These experimental results have shown that data-terms based on the proposed descriptors led to accurate and constant optical flow under varying illumination conditions.