为提高高速公路场景下疲劳驾驶检测的准确率，降低在驾驶员面部易受遮挡、光线复杂的驾驶舱内的疲劳驾驶检测的误判率，提出一种兼顾检测精度与实时性、结合高精度定位与机器视觉的高速公路疲劳驾驶行为检测方法。首先，使用多任务卷积神经网络与FaceNet算法实现驾驶员身份识别，运用改进的实用面部特征点检测器（Practical Facial Landmark Detector, PFLD）算法检测人脸关键点。然后，考虑视频设备工作状态不稳定对疲劳驾驶行为识别产生干扰的场景，基于行驶车辆的轨迹数据，提出纵向位移波动斜率（Slope of Longitudinal Displacement Fluctuation, SLDF），以弥补单一视频设备检测易受光线、遮挡等干扰因素影响的缺陷。随后，使用SLDF指标和3个面部疲劳特征识别疲劳驾驶，并在传统支持向量机（Support Vector Machines, SVM）中加入量子粒子群优化算法，提升SVM分类准确度和缩短运算时间。最后，为验证模型性能进行实车试验，结果表明，复杂场景下疲劳驾驶识别准确率达86.8%，计算时间为3.017 s，与现有其他数据融合算法相比，改进后的SVM分类准确度和运算效率均有提升。融合轨迹、人脸面部多维度、多源的信息有效提高了该系统识别疲劳驾驶的检测精度及其工作的鲁棒性，可为后续高速公路场景下的疲劳特征的检测提供有力支持。

In order to improve the accuracy of fatigue driving detection in highway scenes and reduce the misjudgment rate of fatigue driving detection in cockpit where driver's face is easily obscured and light is complicated, a highway fatigue driving behavior detection method combining detection accuracy and real-time performance, high-precision positioning and machine vision was proposed. Firstly, multi-task convolutional neural network and FaceNet algorithm were used to recognize the driver identity, and the improved PFLD(Practical Facial Landmark Detector) algorithm was used to detect face key points. Secondly, considering the disturbance caused by the unstable working state of video equipment to the recognition of fatigue driving behavior, the SLDF(Slope of Longitudinal Displacement Fluctuation) was proposed based on the track data of running vehicles, to make up for the single video equipment detection is susceptible to interference factors such as light, occlusion. Thirdly, SLDF index and 3 facial fatigue features were used to identify fatigue driving, and quantum particle swarm optimization algorithm was added to traditional SVM(Support Vector Machines) to improve SVM classification accuracy and shorten operation time. Finally, in order to verify the performance of the model, a real car test was conducted, the results showed that the fatigue driving recognition accuracy was 86.8% in complex scenarios, and the calculation time was 3.017 s; compared with other existing data fusion algorithms, the classification accuracy and operation efficiency of the improved SVM were improved. It indicates that the fusion of trajectory, facial multi-dimension and multi-source information effectively improves the detection accuracy and robustness of the system in identifying fatigue driving, and can provide strong support for the subsequent detection of fatigue characteristics in highway scenarios.