基于面部动作和头部姿态疲劳检测方法  

Fatigue Detection Method Based on Facial Movement and Head Posture

作　　者：刘子恒 焦良葆[1] 孟琳 孙宏伟 魏小玉 LIU Ziheng;JIAO Liangbao;MENG Lin;SUN Hongwei;WEI Xiaoyu(Jiangsu Intelligent Sensing Technology and Equipment Engineering Research Center,Institute of Artificial Intelligence Industry Technology,Nanjing Institute of Technology,Nanjing 211167)

机构地区：[1]南京工程学院人工智能产业技术研究院江苏省智能感知技术与装备工程研究中心,南京211167

出　　处：《计算机与数字工程》2025年第3期821-828,共8页Computer & Digital Engineering

基　　金：国家自然科学基金青年基金项目(编号:61903183)资助。

摘　　要：疲劳驾驶是现代交通事故重大成因之一,针对当下疲劳检测方法存在疲劳特征单一、鲁棒性低和眼部疲劳特征难提取的问题,结合面部动作与头部姿态两种维度的视觉信息检测疲劳。该检测方法首先利用MTCNN+PFLD深度网络检测人脸并定位68点人脸特征点,对于面部动作,在对人脸图像进行眼部抠图与数据增强的预处理后,建立深度残差网络对眼部动作级联分类提取眼动信号,同时根据特征点信息计算纵横比,通过设定合理阈值检测嘴部动作;对于头部姿态,采用solvePnP算法结合二维空间内的特征点信息进行坐标系空间变换求解头部姿态欧拉角。最后,根据Perclos算法中的P80原理结合生理实际提取单帧疲劳特征检测结果再进行多帧综合投票,进一步提升鲁棒性。实验结果表明:该检测方法对单帧图像处理平均耗时为49.3 ms,在不同场景下检测精度达97.8%以上,模型泛化能力强,具有较高的实用价值和现实意义。Fatigue driving is one of the major causes of modern traffic accidents.Aiming at the problems of single fatigue feature,low robustness and difficulty in extracting eye fatigue feature in current fatigue detection methods,visual information from two dimensions of facial movement and head posture is combined to detect fatigue.Firstly,MTCNN+PFLD deep network is used to detect the face and locate 68 facial feature points.For facial movements,after eye matting and data enhancement are preprocessed for face images,deep residual network is established to extract eye movement signals by cascade classification.Meanwhile,aspect ratio is calculated according to the information of feature points.By setting a reasonable threshold value to detect the mouth movement.For the head pose,solvePnP algorithm combined with the information of feature points in the two-dimensional space is used to transform the coordinate system space to solve the Euler Angle of the head pose.Finally,according to the P80 principle of Perclos algorithm and the physiological reality,the detection results of single frame fatigue feature are extracted and then multi-frame comprehensive voting is conducted to further improve the robustness.The experimental results show that the average processing time of single frame image is 49.3 ms,and the detection accuracy is more than 97.8%in different scenes.The model generalization ability is strong,and it has high practical value and practical significance.

关 键 词：疲劳驾驶 深度学习 MTCNN PFLD solvePnP 多帧综合投票 PERCLOS 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]