改进YOLOv4的安全帽佩戴检测及身份识别方法  

Helmet Wearing Detection and Identity Recognition Method Based on Improve YOLOv4′s

作　　者：吴冬梅[1] 闫宗亮 宋婉莹 白凡 WU Dong-mei;YAN Zong-liang;SONG Wan-ying;BAI Fan(School of Communication and Information Engineering,Xi'an University of Science and Technology,Xi'anShanxi 710054,China)

机构地区：[1]西安科技大学通信与信息工程学院,陕西西安710054

出　　处：《计算机仿真》2022年第12期290-293,377,共5页Computer Simulation

基　　金：国家自然科学基金-青年项目(61901358);中国博士后科学基金面上项目(2020M673347)。

摘　　要：为了提高建筑施工的效率并降低建筑业事故死亡率,提出一种改进YOLOv4的安全帽佩戴检测及身份识别方法,用于自动监测施工人员是否佩戴安全帽并识别佩戴者的身份。将YOLOv4中的NMS算法改进为Soft-NMS算法;增加主干特征提取网络CSPDarknet53中的残差块个数来提高模型的检测精度;利用迁移学习的方法在四个安全帽数据集上训练,扩大模型的应用范围。实验结果表明,上述算法平均检测精度mAP(Mean Average Precision)达到了92.05%,比YOLOv3、YOLOv4和YOLOv5s高出4.18%、1.66和1.85%,通过对施工现场入口处和施工现场的监控视频进行测试,验证了所提算法在施工现场应用中的有效性。In order to improve the efficiency of construction and reduce accident mortality in the construction industry, an improved YOLOv4 hard hat wearing detection and identification method is proposed for automatic monitoring of whether construction personnel are wearing hard hats and identification of the wearer. First, the NMS algorithm in YOLOv4 was improved to Soft-NMS algorithm, then the number of residual blocks in the network CSPDarknet 53 was increased to improve the detection accuracy of the model, and finally, the application range of the model was expanded by training on the four hard hat data sets by the migration learning method. The experimental results show that the average detection accuracy of this algorithm mAP(Mean Average Precision) has reached 92.05%,which is 4.18%,1.66 and 1.85% higher than YOLOv3,YOLOv4 and YOLOv5s.

关 键 词：安全帽检测 迁移学习 身份识别 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]