Remote gaze estimation based on 3D face structure and iris centers under natural light

摘要

Remote gaze estimation under natural light is still a challenging problem. Appearance based methods are seriously sensitive to illumination variation in the visual spectrum and usually can hardly handle the problem of head movements. And most existing feature-based gaze estimation methods strongly rely on cornea reflections, which are unstable to glasses, head movements and especially useless for natural light condition. In this paper, we propose a novel feature based gaze estimation method without use of cornea reflections. A stereo camera system is built for the proposed method. Firstly, 3D Active Shape Models (ASM) is reconstructed using stereo vision to represent 3D face structure. Then, without use of cornea reflections, a 3D Iris-Eye-Contours based descriptor is proposed to represent human gaze information. Iris centers are used in natural light just like the pupil centers in condition of near-infrared light. What's more, precise estimation of head poses based on 3D face structure is employed to rectify the 3D iris centers and eye contours for improving the ability of tolerance to head movements. Experiments on several subjects show that the system is accurate and allows natural head movements under natural light.