Eye-Model-Based Gaze Estimation by RGB-D Camera

Shigang Li, Jianfeng Li

Abstract: This paper proposes a method of eye-model-based gaze estimation by RGB-D camera, Kinect sensor. Different from other methods, our method sets up a model to calibrate the eyeball center by gazing at a target in 3D space, not predefined. And then by detecting the pupil center, we can estimate the gaze direction. To achieve this algorithm, we first build a head model relying on Kinect sensor, then obtaining the 3D information of pupil center. As we need to know the eyeball center position in head model, we do a calibration by designing a target to gaze. Because the ray from eyeball center to target and the ray from eyeball center to pupil center should meet a relationship, we can have an equation to solve the real eyeball center position. After calibration, we can have a gaze estimation automatically at any time. Our method allows free head motion and it only needs a simple device, finally it also can run automatically in real-time. Experiments show that our method performs well and still has a room for improvement.