Human Eye Detection And Tracking Under Complex Light And Posture

Large amount of data show that driver fatigue is an important cause of combination-unit traffic crashes. For reducing the accident in this aspect, the research in monitoring technology to drivers comes more and more and concentrates on three aspects: the monitoring method based on driver's individual characteristics; the monitoring method based on driver's physiologic parameters, and the monitoring method based on vehicle parameters. The last two methods requires a little wire or the electrode touching the driver's body, not being welcomed by the drivers, or being limited by vehicle type or driving experiences and condition. The first method is a non-contact measuring method. The newest research direction is non-contact measuring method based on perclose.Because of the complex environment the drivers in, the designed system of driver fatigue detection must satisfies that detecting the two eyes very well under the environment of complex illumination, complex posture and people with glasses or not.. After comparing same detection methods, this paper combined the method of support vector machine which based on statistical learning theory with eye detection technology to complete eye detection.This paper mainly completed the eye detection and tracing in driver fatigue detection system. Eye detection is the key technology of the system, first of all, using gray level histogram equalization to eye and non-eye samples which come from different people under complex environment. Represented the result to vector form, then used sequential minimal optimization algorithm to off-line training, accepted a set of support vector and corresponding weight. On the process of on-line detection, firstly, combine background difference method with integral projection method to detection the upper half face, and to reduction the range of eye searching. Then detect eye in the range of the upper face using the result of SVM, completing eye primary detection, finally completing the eye detection using the window fusion method.From the SVM it self can see that this method's speed is very slowly, and can't satisfy the requirement of real-time. Therefore, from the second frame after correct detection, combining kalman prediction and SVM to reduce detection range and improve the algorithm speed.Experiments showed that this method can realtime complete eye detection basically, and have same adaptability to complex illumination, complex posture and people with glasses or not.