Research On Driver Fatigue Detection Using Depth Information And Color Image

With the rapid development of national economy, the purchasing power of people increasing rapidly, and are no longer strange to cars which are used for traveling in life. But with the coming vehicle problems it draws more attentions, traffic accidents and other problems have becoming increasingly prominent. With so many cars accidents occurring frequently, in the final analysis human factors account for most of them. The drivers continue driving while they feel tired, leading to tragedy. So it is particularly necessary to monitor drivers' driving state, when driver fatigue occurs it alarms in time.We summarize the general methods of driver fatigue detection, and determining to use more mainstream approach which is based on computer vision. As Kinect sensor can both get depth information and color image at the same time, using these two to detect head pose and eye state so to judge fatigue driving. The main research work are as follows:Firstly, after getting color image stream from Kinect sensor, detecting the eye state from the image. Firstly, doing face detection using AdaBoost algorithm. In order to accelerate detection speed, we can appropriately increase the minimum search window and zoom factor during face detection. Upon complication of the human face detection, using facial features prior knowledge of layout to divide human eye area roughly, and then do median filtering and binary processing. Then the next mission is about precise positioning of the human eye, mainly using histogram integral project to locate the center position of human eye, and getting the minimum rectangle area of the orbit according to the center position. We set three parameters weighted evaluation based on the minimum eye rectangle, obtaining the eye state through fuzzy combined evaluation method.Secondly, as the dark light situations, eye state detection result is not so good even complete failure. Because depth information is not affected by illumination, we use discriminative random regression forests to do real-time head pose estimation. Firstly, we introduce the principle of decision tree construction algorithm, based on the algorithm leading us to random forests. Then, based on traditional random regression forests, we join the method of classification since the depth image we obtain including not only the head portion, as well as the human body part. Under different lighting conditions, we do head pose estimation experiments while driver shows different poses or facial expression and wearing different clothes. Test results show our method is robust.Finally, we collected the real-world driving videos to judge fatigue driving. We use PERCLOS rule to judge eye tired and at the same time we use head moving angles to judge doze. The results show that combination of these two ways to determine fatigue driving is more robust and accuracy.