Design And Implementation System For Driver’s Head Posture Estimation

With the rapid development of world economy, the transportation industry followed aswell, make cars been one of indispensable traffic tools in people’s lives. However, thetraffic accidents increasing inescapability as the number of cars increase. One of aimportant reason of traffic accidents is fatigue driving. According to statistics, fatiguedriving is in the third place in the risk factors of traffic accidents, If the monitoring of thedriver is build to give a warning in time when the driver is tired, the traffic accidentscaused by this reason will reduced greatly, Give an accurate estimate to the driver’s headposture is an important part in the driver’s fatigue detection.Usually, there are two kinds of method for head orientation estimation, model-basedand appearance-based. The method based on the model is easy to understand, simple toimplement and little for calculate, the method based on the appearance is not sensitive tothe head posture and have high robustness. Combine with the advantage of the twomethods can get an accurate result of the driver’s head posture estimation. The methodbased on the model locate the whole face and then locate the feature points of the face,eyes and mouth, then form a characteristic triangle which will be changed with therotation of the driver’s head. Through find the rule to estimate the driver’s head postureaccurately. But this method is not suitable for driver’s head with a wide range because itcan’t detect the feature points of the face at this moment. Then the method based on theBP neural network can be used to estimate the driver’s head posture. Its main idea isthrough supervised learning and training to get the mapping relationship between the headposture and a2D image. At first, confirm the network structure according to the practicalproblems, the number of neurons in the input layer and the output layer are easy todetermine, the most appropriate value of the number of neurons in the hidden layer will beget in the training. Deal with the training sample to get expected value of the input dataand expected value of output data of the network, stop the network training and save theweight value when the allowed error is achieved. Then can be used to estimate the headposture of the driver.Experiments show that combined with the advantage of the two methods can effectively improve the real-time performance and robustness of the driver’s head postureestimation.