Direct visual tracking under extreme illumination variations using the sum of conditional variance

Abstract

Gradient-based optimization is a very efficient strategy to solve the direct visual tracking (DVT) problem using transformation models with many degrees of freedom (DOF). Even though popular DVT methods use the sum of squared differences as similarity function, this approach is not robust to illumination variations often verified in practice. One technique to compensate illumination variations is through an illumination model, which, in turn, increases the total number of parameters to be computed. High quality augmented reality and robotic systems demand fast tracking speeds, which can be impaired by the computational complexity added by the illumination model. In this paper, we propose a robust DVT method capable of tracking in extreme illumination conditions. Building upon the sum of conditional variance as similarity function, we propose a novel tracking method that significantly reduces the computational effort compared to similar methods proposed in the literature. We provide extensive experiments and quantitative analysis using challenging videos to attest the advantages of the proposed method.