赵彦龙1  赵三伟2  闫伟才3  钟震宇4

（1.内蒙古军区数据信息室，内蒙古 呼和浩特 010051

2.武汉滨湖电子有限责任公司，湖北 武汉 430205

3.中通服咨询设计研究院有限公司，江苏 南京 210023

4.广东省科学院智能制造研究所/广东省现代控制技术重点实验室，广东 广州 510070）

摘要：在生产作业过程中，作业人员注意力不集中是生产事故发生的主要诱因。针对现有的专注性研究方法过于依赖高性能计算设备，存在部署困难和隐私数据泄露等问题，提出一种结合深度学习和边缘计算的作业人员专注性检测方法。首先，利用YOLOv5算法对使用手机、抽烟和喝水3种常见的分心行为进行实时检测；然后，采用轻量化人脸关键点检测网络和Perclos算法对作业人员的疲劳程度进行评估，并对疲劳状态进行预警；最后，采用基于通道剪枝的压缩算法将分心行为检测网络和疲劳行为检测网络部署于低功耗的便携式边缘计算设备，避免作业人员隐私和商业数据泄露的风险。实验结果表明，该方法对疲劳行为和分心行为的检测准确率分别达98.6%和99.8%，满足实际的部署需求。

关键词：专注性检测；YOLOv5算法；人脸关键点检测网络；Perclos算法；通道剪枝的压缩算法

中图分类号：TP391.4           文献标志码：A            文章编号：1674-2605(2023)05-0007-06

DOI：10.3969/j.issn.1674-2605.2023.05.007

Detection Method of Operator Concentration Combined with             Deep Learning and Edge Computing 

ZHAO Yanlong1  ZHAO Sanwei2  YAN Weicai3  ZHONG Zhenyu4

(1.Data Information Office, Inner Mongolia Military Region, Hohhot 010051, China

2.Wuhan Binhu Electronics Co., Ltd., Wuhan 430205, China

3.China Comservice Consulting Design Research Institute Co., Ltd., Nanjing 210023, China

4.Institute of Intelligent Manufacturing, Guangdong Academy of Sciences/Guangdong Key Laboratory 

of Modern Control Technology, Guangzhou 510070, China) 

Abstract: During the production operation process, the lack of concentration of operators is the main cause of production accidents. In view of the existing focus research methods that rely too much on high-performance computing devices, which have problems such as deployment difficulties and privacy data leakage, a focus detection method for operators combining deep learning and edge computing is proposed. Firstly, the YOLOv5 algorithm is used to detect three common distracting behaviors: using a mobile phone, smoking, and drinking water in real-time; Then, a lightweight facial keypoint detection network and Perclos algorithm are used to evaluate the fatigue level of operators and provide early warning of fatigue status; Finally, the channel pruning based compression algorithm is used to deploy the distraction behavior detection network and fatigue behavior detection network to low-power portable edge computing devices to avoid the risk of operators' privacy and commercial data leakage. The experimental results show that the detection accuracy of this method for fatigue behavior and distraction behavior reaches 98.6% and 99.8%, respectively, meeting the actual deployment requirements.

Keywords: concentration detection; YOLOv5 algorithm; face key point detection network; Perclos algorithm; compression algorithm for channel pruning