Design And Implementation Of Driver’s Head Posture Estimation

With the economy and society developing, driving car has become one of the mostimportant choices for outside. At the same time, traffic accidents take place in a rapid rateinevitably. Fatigue driving is one of the major risks for the road safety nowadays, and itleads to accident and results in personal injury and property lost. According to statistics,the casualties caused by fatigue driving are not less than drunken driving. Usuallyabnormal head movements occur when the driver is fatigue. Therefore, a real-timemonitoring of the driver’s head and accurately estimates the abnormal head posture is avery important part in the detection of fatigue driving.Model-based estimation is simple, easy to understand and calculation smaller,appearance-based estimation is robust and suitable for low-resolution images. Thisthesis takes the advantage of both, complements driver’s head posture estimation.Model-based method Firstly locates the position of eyes and mouth, and thenconstructs an isosceles triangle. Based on the change of isosceles triangle, it estimatesthe posture. In order to estimate a wide range of head change, the first step is extractsthe gradient feature of the image, the gradient feature can effectively solve the problemcaused by light. However, the gradient feature of image is merely a first-orderstatistical feature. If dividing image to be many grid, we find out that adjacent imagegrids are not entirely independent, there is some relationship between, and thisrelationship is different for different head posture. Therefore the second steps wefurther extract the gridded Image Euclidean Distance feature, firstly we divide thegradient feature image into many grids, suitable dividing size can reduce featuredimension and result in reducing the amount of calculation. Then calculate the imagegrid Euclidean distance with each other, the results of the calculation create a newposture descriptive feature. The gridded image Euclidean distance feature can depictthe relationship between adjacent image grids clearly, the feature contains a wealth ofinformation, and is closely related to head posture, and therefore it possesses a moredescriptive ability and posture distinguishing ability. The third step we carry out PCAdimension reducing. At the end, we classify the posture features by adaptive classification.Experimental result shows that it improves the validity and accuracy of driver’s headposture estimation.