Research On Detecting Method Of Driver's Visual Attention

Driver's fatigue and visual inattention are the main factors of traffic accidents. Therefore, developing systems for monitoring the driver's level of visual attention and alerting the driver accurately is necessary to prevent accidents. As the development of computer vision technology, detecting the driver's eyes and mouth in real-time and determining the driver's level of visual attention have become the main researches on monitoring driver's fatigue and visual attention.The thesis describes a real-time system based on computer vision technology for monitoring driver's facial features including positoions of face, eyes, nose and mouth corners. The system starts with acquiring a piece of consecutive video images of the driver's face using a CCD microcamera. And the driver's eyes and mouth corners in every frame are positioned by the fast and compact algorithm. Then the driver's level of visual attention is determined by analyzing the information of feature points in images.The major work of the thesis focuses on the algorithm research of detecting system of driver's visual attention. First, the driver's face in the first frame is detected by using skin color detecting. After selecting the nose region according to the geometrical relation of face feature points, a new methodology for positioning nose feature point based on BP neural network and shape from shading (SFS) is proposed. Then the system positions eye corners and mouth corners quickly around the nose point by using mouth color detecting and improved horizontal Sobel edge detecting. The system positions the eye corners and mouth corners for the images except the first frame directly based on template matching. Finally, an original rule of determining the driver's level of visual attention is presented. According to the rule, the system obtains the sign number of the driver's level of visual attention in every frame and determines the driver's level of visual attention integrating the sign numbers of previous images. Experimental results demonstrate that the proposed detecting algorithm appears to be simple, reliable and robust.