Research Of Gaze Tracking Technology Based On Image Processing

It is always an important direction to improve the controllability of machine in thescientific research of HMI (human-computer interaction) filed. Voice recognition andgesture recognition technology pioneered the HMI application filed, followed by therapidly developing eye tracking technology, we can imagine its new interactive experiencewith those novel technologies successfully implied in the actual project. Applying videoimage processing technology in system we accomplish real-time eye tracking and gazedirection tracking. We use a CCD camera to capture video sequences under thenear-infrared light illumination, then enable eye tracking algorithm to track and extract eyeregion image, finally obtain the relevant parameters to calculate the sight placementcoordinates. After going through the pre-existing test and analyzing the pros and cons of avariety of methods, we focus on the eye tracking and gaze point estimation technology, andmake innovation in eye detection, target tracking, estimation of gazing point andcompensation algorithm for the illumination and head offset. The main work in this thesis isas follows:1. An acute problem in the eye detection system is the high recognition accuracy.Besides, we use the detection outcome as a basis for system initialization. In order toguarantee the recognition accuracy, we use eye detection classifier trained by improvedAdaboost algorithm to obtain initial information. In the improved algorithm, we add samplestatus flags and some negative samples under near-infrared light illumination. By theanalysis of experiment, the improved classifier show high accuracy in eye detection andlocation, and also adapt to the near-infrared condition.2. Eye tracking system show high demand for real-time ability. Considering theshortage of light and dark pupil image difference method, and the dark image show a goodfeature in pupil contour, easy to implement pupil region segmentation, we decide to capturethe dark pupil for image processing based on particle filter algorithm. The particle filteralgorithm with its observed model of grayscale image histogram features and imagevariance characteristics is used to track the human eye area. While feature histogramexplain image from statistical level, the variance explain image from grayscale distributionstructure. Those two complementarity feature achieve image matching between the videosequences. By the analysis of experiment, this algorithm not only can be as precise as the detection part, but also reduce the processing time, improve instantaneity of system.3. Taking into account the unlimited head offset may impact the changing of the eyeposition and bring highlight illumination under actual environment, we designedcompensation algorithm corresponding both depth and parallel direction offset. In addition,in order to improve the adaptive ability in the dark pupil contour segmentation, a robustnonlinear illumination compensation algorithm is designed to fit the face illuminationchange caused by head movement. The experiment shows that the compensation algorithmsenhances the stability of system and improve precision in gazing point estimation part.Finally, we built demo platform for gaze tracking system, and test the whole system.Test result show the system performs well in recognition accuracy, instantaneity andnumerical accuracy.