Research studies on machine vision-based driver fatigue detection algorithm have improved traffic safety significantly. Generally, many algorithms asses the driving state according to limited video frames, thus resulting in some inaccuracy. We propose a real-time detection algorithm involved in information entropy. Particularly, this algorithm relies on the analysis of sufficient consecutive video frames. First, we introduce an improved YOLOv3-tiny convolutional neural network to capture the facial regions under complex driving conditions, eliminating the inaccuracy and affections caused by artificial feature extraction. Second, we construct a geometric area called Face Feature Triangle (FFT) based on the application of the Dlib toolkit as well as the landmarks and the coordinates of the facial regions; then we create a Face Feature Vector (FFV), which contains all the information of the area and centroid of each FFT. We use FFV as an indicator to determine whether the driver is in fatigue state. Finally, we design a sliding window to get the facial information entropy. Comparative experiments show that our algorithm performs better than the current ones on both accuracy and real-time performance. In simulated driving applications, the proposed algorithm detects the fatigue state at a speed of over 20 fps with an accuracy of 94.32%.

中文翻译：

---

基于面部运动信息熵的疲劳驾驶检测算法

基于机器视觉的驾驶员疲劳检测算法的研究显着提高了交通安全性。通常，许多算法根据有限的视频帧来评估驱动状态，从而导致一些不准确。我们提出一种涉及信息熵的实时检测算法。特别地，该算法依赖于足够的连续视频帧的分析。首先，我们引入了一种改进的YOLOv3-tiny卷积神经网络，以捕获复杂驾驶条件下的面部区域，从而消除了人工特征提取所引起的不准确性和影响。其次，基于Dlib工具包的应用以及面部区域的地标和坐标，我们构造了一个称为“面部特征三角（FFT）”的几何区域；然后我们创建一个人脸特征向量（FFV），其中包含每个FFT的面积和质心的所有信息。我们使用FFV作为指标来确定驾驶员是否处于疲劳状态。最后，我们设计一个滑动窗口以获取面部信息熵。对比实验表明，我们的算法在准确性和实时性能上都比当前算法更好。在模拟驾驶应用中，提出的算法以超过20 fps的速度检测疲劳状态，精度为94.32％。