Research On Eye-state Based Monitoring For Drivers' Dozing

Recently, driving wearily has been the main reason in car accident. As we all know, fatigue always brings doze. The problem is drivers often do not know they have already got into a dozing and dangerous situation. Therefore, we need to design a method to monitor the drivers' attention at the real-time level.The main parameters to monitor doze are drivers' physiology signal, cars' running parameters and fatigue strength parameters for mathematic modeling. However, it is always unreliable or hard to monitor with all the parameters above mentioned. Some even needs to use additional devices which are uncomfortable for drivers for a long time. So the vision based doze monitoring method has been the mainstream for its characteristic of non-touch and reliability.The thesis proposes an eye-state based doze monitoring method which is a vision based method. In the thesis, two aspects are improved compared with the traditional one that is face detection and doze recognition. We propose the new face detection method basing on the AdaBoost method, which is sample-learning method. As a turning point to accelerate face detection, it is really a practical and useful method, but it can not detect the faces with various poses. So a new multi-pose oriented face detection algorithm is proposed, which is well adapted to solve this problem. As to the doze recognition, the eye-state identification is the most important part. Compared with the traditional method, the thesis proposes a new algorithm, in which two eye features are quantified and fused to make a more accurate detecting result. The curvature of the upper eyelid and eye's area are selected as the fusion features ultimately.Finally, based on the two improved strategies mentioned above, the thesis proposes a doze monitoring solution. The experiments to the solution show its excellence both in real-time and accuracy performance.