Dual-Mode Pupil and Iris Gaze Tracking Systems for Real-Time Wearable Eye Trackers

• Main Authors: Jia-Hao Wu, 吳家豪
• Other Authors: Chih-Peng Fan
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/64851290203988761120

碩士 === 國立中興大學 === 電機工程學系所 === 104 === Recently, wearable devices provide an easy and convenient human-computer interaction in our life directly. In this thesis, a low cost wearable eye tracker system is implemented. After the user completes the calibration process, the proposed system allows head free and estimates the gaze points simultaneously and correctly. In the thesis, we propose the dual-mode pupil and iris gaze tracking systems which use one camera to capture the eye image and use the other camera to capture the scene image. The infrared based camera for eye images can segment pupil and sclera information, or the user can turn off the infrared function and change to visible light mode to capture the iris manually. Then the system changes to ellipse fitting and estimates the pupil and iris centers. In addition, the scene camera captures image from user''s view as the eye of user. The proposed algorithm is divided into: capture eye images, search ROI of pupil (or iris), pupil (or iris) segmentation, ellipse fitting with the RANSAC based process, and catch calibration points for calculating gaze placement. In our experiment, we utilize 9 calibration points to increase the accuracy compared with 4 calibration points. The system performs some distortions when the user gazes the corner points of square boundary. Since the human eyes are stereo-sphere shape, the coordinates cannot be a square when the gaze points project from 3-dimensions space to a 2-dimensions image. By implementing on the PC platform with Intel i7-3770 at 3.4GHz operational frequency, our experimental results show that the offset of pupil center are both within 2 pixels for horizontal and vertical coordinates. By using visible light for tracking iris technology, its center offset of horizontal is 7 pixels, and the vertical offset is 5 pixels. Through the perspective function, the system achieves the accuracy between 0.1° and 1.46° for horizontal coordinates, and 1.09° and 2.68° for vertical coordinates for the gaze placement error. Besides, the system achieves the accuracy between 0.85° and 1.51° for horizontal coordinates, and 1.53° and 3.56° for vertical coordinates for the gaze placement error.