改进多任务级联卷积神经网络的驾驶员疲劳检测  

Improved Multi-task Cascaded Convolutional Networks for Driver Fatigue

作　　者：刘星 文良华 成奎 陈波杰 张宇杰 于剑桥 LIU Xing;WEN Lianghua;CHENG Kui;CHEN Bojie;ZHANG Yujie;YU Jianqiao(School of Electronic Information and Engineering,Yibin University,Yibin 644000,China;School of Computer Science and Technol-ogy,Yibin University,Yibin 644000,China;Sanjiang Artificial Intelligence and Robot Research Institute,Yibin University,Yibin 644000,China;Intelligent Terminal Industry College,Chengdu Technological University,Yibin 644000,China)

机构地区：[1]宜宾学院电子信息工程学院,四川宜宾644000 [2]宜宾学院计算机科学与技术学院,四川宜宾644000 [3]宜宾学院三江人工智能与机器人研究院,四川宜宾644000 [4]成都工业学院智能终端产业学院,四川宜宾644000

出　　处：《宜宾学院学报》2024年第12期7-11,68,共6页Journal of Yibin University

基　　金：四川省科技计划项目(2023YFQ0093);宜宾学院培育项目(2023PY10)。

摘　　要：针对驾驶员疲劳检测方法中存在单一特征检测的局限性,且由于模型参数计算量过大导致在低算力的移动边缘计算设备上检测耗时过长的问题,提出一种改进的多任务级联卷积神经网络(MTCNN).通过对子网络R-Net的优化,采用平均池化来减少模型参数量,并将全连接层替换为均值池化,结合Dlib对人脸64个特征点的精准定位,选取效果较好的阈值参数实现疲劳检测.实验结果显示,在人脸数据集WIDER FACE和LFW数据集上,改进后的算法相比于改进前,参数量减少了47.5%,人脸检测的准确率从96.7%提升至97.8%.最后通过YawDD疲劳数据集,在资源受限的树莓派4B设备上实现了高效的疲劳检测,验证了其在实际应用中的可靠性.Aiming at the limitations of single-feature detection in driver fatigue detection methods and the time-consuming detection on low-computing-power mobile edge computing devices due to the excessive computation of model parameters,an improved multi-task cascaded convolutional neural network(MTCNN)was proposed.Through the optimization of the sub-network R-Net,the average pooling was used to reduce the number of model parameters,and the fully-connected layer was replaced by mean pooling,combined with the accurate localization of the 64 feature points of the face by Dlib,and the threshold parameter with better effect was selected to achieve fatigue detection.The experimental results show that on the face dataset WIDER FACE and LFW dataset,the improved algorithm reduces the amount of parameters by 47.5%compared to the pre-improvement one,and the accuracy of face detection increases from 96.7%to 97.8%.Finally,through the YawDD fatigue dataset,efficient fatigue detection is realized on the resource-constrained Raspberry Pi 4B device,which verifies its reliability in practical applications.

关 键 词：深度学习 疲劳检测 MTCNN 树莓派 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]