Gaze Tracking And Its Application In Driver And Passenger Monitoring System

Gaze tracking is a technique to obtain the focus of people by making use of mechanical, electronic, optical and other methods, which is widely used in human computer interaction, automobile auxiliary driving, military, human behavior study and disease diagnosis. Gaze tracking can be classified into two different types called the intrusive and nonintrusive also known as wearable and telemetry according to different system hardware. With the development of computer technology, pattern recognition and other technology, gaze tracking based on analysis of digital video (Video Oculographic, VOG) is becoming an active research topic.In VOG system, computer vision is used to get human face images and detect eye features. The gaze parameters can be extracted from these features and then by using them the gaze direction or gaze point can be obtained. Thus, VOG gaze tracking system is composed of two components:gaze feature parameter extraction and gaze direction estimation.A single camera gaze tracking system is developed in this dissertation to study key technologies on gaze tracking. The research work in this dissertation is composed of four parts, namely location of human face and human eye, gaze feature detection, mapping model establishment between gaze parameters and gaze points on screen, gaze points compensation under the condition of head movement. In the final part, we also discuss the application of gaze tracking in driving assistance. The innovation of our research is as follows.In the stage of human face and eye location, the method which combines skin color model with Adaboost is used to locate human face. Skin color model can get good and fast face detection on condition that the target and background are obviously different in the image, which is insensitive for face gesture, but false detection rate is too high. Adaboost method has good performance for detecting the positive face and the false detection rate is low, which is not suitable for multi-pose face. This paper combines the advantages of both by rotating the image by a certain angle to detect the respective angles of the face, in order to achieve multi-pose face detection and improve the detection speed. After face location eye region in the face can be determined in accordance with prior knowledge of the facial distribution. In this eye region Adaboost is used again to locate human eye and the extend area of iris is obtained. In the following steps eye feature extraction is implemented in the located eye region, which can obtain gaze parameters, namely the vector from the canthus of eye to the center of iris.After gaze parameters is obtained the nonlinear mapping model between gaze parameters and gaze points need to be established. In the existing paper nonlinear polynomial is often used as the mapping model to describe the relationship between gaze parameters and gaze points. Mapping model parameters are obtained through setting the calibration points on the screen, and calibrate the polynomial model parameters. In this paper an ELM based mapping model is proposed for its advantages of suitable for multi classification and also have fast training and strong generalization ability. In ELM based method gaze parameters is served as the network input, gaze points on the screen as output, and nonlinear polynomial as activation function. The training samples are obtained by initial calibration, so mapping model between gaze parameters and gaze points is established. The experimental results show that ELM based method is better than classical nonlinear polynomial model method not only in accuracy but also in stability and is a new way to achieve the mapping model in single camera gaze tracking system.Most of current gaze tracking methods are under the condition of head stationary, particularly single camera based gaze tracking system, to calibrate the mapping model between gaze parameters and gaze points. So it is better to keep head still and stay a certain distance from the screen. For this kind of system when head position is changed compared with the calibration position, the gaze points cannot be estimated accurately. Based on single camera system without light source we have proposed a method of gaze point compensation under head movement through the camera model and coordinate transformation. Experimental results show that the compensation method has excellent performance in real application. The method can also be applied in other types of gaze tracking system.