基于深度学习的司机疲劳驾驶检测方法研究

Research on Driver Fatigue Driving Detection Method Based on Deep Learning

针对传统基于机器视觉的司机疲劳检测模型对硬件系统要求较高、检测准确率和效率较低等问题,提出一种基于MTCNN-PFLD-LSTM深度学习模型的疲劳驾驶检测算法。通过多任务卷积神经网络MTCNN进行人脸区域检测;利用PFLD模型检测人脸眼部、嘴部和头部的关键点及空间姿态角;计算出基于时间序列的人脸疲劳特征参数矩阵并输入长短期记忆网络LSTM进行疲劳驾驶检测,通过优化设计不同阶段损失函数及其权重,进一步提高检测能力。在未采用GPU加速的情况下,通过YawDD数据集与自采数据集进行试验并与最新的8种方法进行比较,准确率和检测帧率分别达到99.22%和46,准确率比未采用GPU加速试验中性能第2的模型增加了0.26%,检测帧率比未采用GPU加速试验中性能第2的模型增加了1.3倍。试验结果表明,提出的方法可以提高疲劳检测的准确度和效率,并可在移动设备等低算力设备上应用。

In response to the problems of the traditional driver fatigue detection algorithms based on machine vision,such as high requirements for computer hardware,low accuracy and efficiency,the paper proposed a fatigue driving detection algorithm based on MTCNN-PFLD-LSTM deep learning model.First,the face contour was detected by multi-task convolutional neural network(MTCNN).Second,the praclical facial landmark detector(PFLD)algorithm was used to detect the key points of the eyes,mouth and head,as well as the spatial attitude angles.Finally,the face fatigue parameter matrices based on time series were calculated and input into the long short-term memory network(LSTM)for fatigue driving detection to further improve the detection ability by optimizing the loss function and its weights in different procedures.The experiments were done on the YawDD dataset and self-collected dataset without using GPU acceleration and compared with the latest 8 algorithms.The accuracy rate was up to 99.22%,which is increased 0.26%than the second best algorithm without using GPU acceleration.The frames per second(FPS)were up to 46,which were increased 1.3 times than the second best algorithm without using GPU acceleration.The results show that the algorithm proposed in this paper has improved accuracy and efficiency of fatigue driving detection,and can be applied to devices with low computing power,such as mobile devices.