Robust Non-Intrusive Eye Tracking And Gaze Estimation

As an important part of human body, eyes can not only help us explore the unknown world but also is a window of human heart. Robust eye tracking and gaze estimation technology is being applied into many areas such as psychology, military, marketing management, computer vision, medical science and so on. And gaze estimation based human machine interaction is becoming more and more popular and also brings new challenges to the traditional way of human machine interaction. And the non-intrusive eye tracking and gaze estimation system receives the highest attention from the researchers. This paper reviews recently development of in the area of eye tracking and gaze estimation and proposes a fast and accurate eye center location method and also an eye center tracking method based on the former eye location method. Furthermore this paper also proposed a gaze estimation method that allows large scale of head movement.1. In order to deal with the accuracy and speed of eye center location problem, this paper proposes a fast and accurate eye center location method. Firstly the face area in the image is detected and the rough eye regions are extracted through anthropometric relations. Then different kinds of kernels are designed to convolution the extracted eye regions. The final eye center is determined by calculating the maximum value in the division matrix of two kinds of convoluted eye area images. By using the FFT to calculate the convolution, the proposed method can achieve fast performance. And the affection of illuminations, occlusion of eyelid and glasses are reduced by considering the gray value of eye center pixel. The proposed method is tested on a public available database Bio ID, and the result shows that the proposed method outperforms the state of the art eye location methods in terms of accurate and computing cost.2. Most of the current gaze estimation methods use high resolution eye image or an assistant IR light to calculate the gaze direction or fixing point in a screen. This paper proposed a gaze estimation method that can be applied to normal web camera without the need of IR light source. Firstly the Supervised Descent Method is used to calculate the facial features of the detected face image. Secondly the POSIT algorithm is used to determine the head pose. This paper build an eye feature using eye corners, eyelids, and iris centers. Then the adaptive linear regression is used to map the detected eye features to three angles of the gaze when head is still. The final gaze direction is determined by adding the three angles to the head pose. Thus the proposed gaze estimation system can handle large head movements.Finally, this paper analyze the theory and also the result of the proposed method. Further improvement is also addressed.