关键词: 卷积神经网络, 光流跟踪, 打哈欠检测, 信息融合, 二比特特征

Abstract: A method for driver yawning detection was proposed based on multi-information fusion. Firstly, the driver face database was built ，and the deep convolutional neural network was trained for driver face and nose detection in sequence. Secondly, an online nose detector was generated through training a random forest classifier with local binary feature,which was introduced to compensate the nose tracking errors caused by optical flow-based nose tracker. At last, the edge gradient changes were analyzed in the mouth area under the nose. The informations like nose tracking confidence and face lateral movement were combined to determine the drivers were yawning. The experimental results show that, compared with other face classification method, the deep convolution network may obtain better classification detection effectiveness. The online learning-based tracking method may reduce the drift errors greatly introduced by the optical flow-based tracker. The whole algorithm may detect driver yawning event with satisfactory accuracy under different driving conditions.

Key words: convolutional neural network, optical flow tracker, yawning detection, information fusion, binary feature