Reseach And Implementation On Non-intrusive Eye Tracking

Along with the development of computer technology and nonlinear filtering theory, eye detection and tracking has become one of the hotspots in the fields of computer vision and automation. The present eye detection and tracking methods were studied and analyzed in this dissertation, and their deficiencies were improved.Aiming at the problem that the classifier based on adaboost algorithm was time-consuming in the training stage and prone to over-training, optimizing haar features, enriching training samples and automatic combining cascade classifier and other measures were adopted in this dissertation to improve training speed. In the detection stage, in order to improve the detection accuracy and shorten the detection time, the image was preprocessed by gray-scale, histogram equalization and median filtering, and the detection results was supervised by principal components analysis. According to the inherent characteristics of iris, susan operator was used to eliminate the influence of eyeball-like factors, such as eye brows, then connected-component analysis was employed to pinpoint iris. Allowing for the prior knowledge about iris’ s specific ellipticity, in the initializing tracking template stage, the center and the size of the template was adjusted to more precisely. For traditional camshift algorithm, noise suppression was a weak link. To effectively eliminate the impact of noise, the characteristics of the noise was analyzed in this dissertation and the relation between the minimum saturation and the size of search window was established to adaptively adjust screening threshold in the back projection stage. To accurately track iris, the gradient vector feature was adopted on the basis of color feature to improve multi-features-fusion method and enhance the difference between iris and distractions. To obtain the motion state of eyes, an algorithm based on kalman filter and improved camshift was proposed in this dissertation. The position of search window located by camshift algorithm was used as the observations of kalman filter to modify the estimated value and act as the criteria of camshift algorithm in next frame to improve tracking accuracy. Taking into account the situation of iris deformation and non-uniform motion, such as strabismus and closed eyes, edge feature of the iris was full used by a custom criterion to judge and reforecast iris’ s center on the basis of its historical trajectory. Finally, to determine the line of sight range, a coordinate transformation model from the image coordinate system to the physical world was proposed.Through experimental analysis, these proposed methods in this dissertation have the advantages of faster speed, higher accuracy in the eye detection stage and high real-time, strong robustness, high accuracy in the eye tracking stage under different environment. Even more, they have good performance in determining the line of sight in a certain range and are able to reach the requirement of non-linear eye tracking.