Most current vision-based fatigue detection methods don’t have high-performance and robust face detector. They detect driver fatigue using single detection feature and cannot achieve real-time efficiency on edge computing devices. Aimed at solving these problems, this paper proposes a driver fatigue detection system based on convolutional neural network that can run in real-time on edge computing devices. The system firstly uses the proposed face detection network LittleFace to locate the face and classify the face into two states: small yaw angle state “normal” and large yaw angle state “distract.” Secondly, the speed-optimized SDM algorithm is conducted only in the face region of the “normal” state to deal with the problem that the face alignment accuracy decreases at large angle profile, and the “distract” state is used to detect driver distraction. Finally, feature parameters EAR, MAR and head pitch angle are calculated from the obtained landmarks and used to detect driver fatigue respectively. Comprehensive experiments are conducted to evaluate the proposed system and the results show its practicality and superiority. Our face detection network LittleFace can achieve 88.53% mAP on AFLW test set at 58 FPS on the edge computing device Nvidia Jetson Nano. Evaluation results on YawDD, 300 W, and DriverEyes show the average detection accuracy of the proposed system can reach 89.55%.

当前大多数基于视觉的疲劳检测方法都没有高性能且坚固的面部检测器。他们使用单一检测功能检测驾驶员疲劳，并且无法在边缘计算设备上实现实时效率。为了解决这些问题，本文提出了一种基于卷积神经网络的驾驶员疲劳检测系统，该系统可以在边缘计算设备上实时运行。该系统首先使用所提出的面部检测网络LittleFace定位面部并将面部分类为两种状态：小偏航角状态为“正常”和大偏航角状态为“分散”。其次，仅在“正常”状态的面部区域进行速度优化的SDM算法，以解决在大角度轮廓处面部对齐精度降低的问题，“分心”状态用于检测驾驶员的分心情况。最后，从获得的界标中计算出特征参数EAR，MAR和头部俯仰角，并分别用于检测驾驶员疲劳程度。进行了综合实验，对所提出的系统进行了评估，结果表明了该系统的实用性和优越性。我们的面部检测网络LittleFace在边缘计算设备Nvidia Jetson Nano上以58 FPS进行AFLW测试时，可以达到88.53％的mAP。在YawDD，300 W和DriverEyes上的评估结果表明，该系统的平均检测精度可以达到89.55％。进行了综合实验，对所提出的系统进行了评估，结果表明了该系统的实用性和优越性。我们的面部检测网络LittleFace在边缘计算设备Nvidia Jetson Nano上以58 FPS进行AFLW测试时，可以达到88.53％的mAP。在YawDD，300 W和DriverEyes上的评估结果表明，该系统的平均检测精度可以达到89.55％。进行了综合实验，对所提出的系统进行了评估，结果表明了该系统的实用性和优越性。我们的面部检测网络LittleFace在边缘计算设备Nvidia Jetson Nano上以58 FPS进行AFLW测试时，可以达到88.53％的mAP。在YawDD，300 W和DriverEyes上的评估结果表明，该系统的平均检测精度可以达到89.55％。