INFRARED IMAGE GAZE ESTIMATION BASED ON IMPROVED HOURGLASS NETWORK

With the rise of metauniverse concepts such as VR/AR, gaze estimation technology integrating iris recognition is a research hotspot at present. Both iris recognition and gaze tracking use near-infrared cameras to capture eye images. It is a challenging task to estimate the line of sight only from monocular infrared eye images. In this paper, a new gaze estimation network is proposed. The middle eye model label image was obtained by using the mapping of circular eye and elliptical iris, and the intermediate supervision was introduced. The middle eye model image of the infrared iris image was regressed through the improved hourglass network, and the 3D gaze angle was regressed with DenseNet. The near-infrared rainbow film acquisition system was designed, and the infrared iris dataset SEPAD-Gaze was established. The experimental results show that the error of the gaze estimation method proposed in this paper is 4.62° on the dataset SEPAD-Gaze. The generalization verification experiment is done on the public dataset MPIIGaze under visible light, the error is as low as 4.48°, which exceeds all the current mainstream algorithms.