改进YOLO v4算法的电动车驾驶员头盔佩戴检测  

Helmet Wearing Detection for Electric Vehicle Drivers Based on Improved YOLO v4 Algorithm

作　　者：吴冬梅[1] 尹以鹏 宋婉莹 王静[1] WU Dong-mei;YIN Yi-peng;SONG Wan-ying;WANG Jing(College of Communication and Information Engineering,Xi'an University of Science and Technology,Xi'an,Shanxi 710600,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710600

出　　处：《计算机仿真》2023年第3期508-513,共6页Computer Simulation

基　　金：国家自然科学基金-青年项目(61901358);陕西省教育厅科研计划项目(19JK0528);陕西省教育厅一般专项项目(20JK0757)。

摘　　要：针对电动车驾驶人员未佩戴头盔的现象,提出了一种改进YOLOv4(You only look once)算法的电动车驾驶人员头盔佩戴检测方法。将数据集利用K-means算法进行聚类以获得先验框,提高先验框与特征图的匹配程度;在CSPDarknet53主干特征提取网络的输出层增加卷积层,并在PANet(Path Aggregation Network)网络部分增加SPP(Spatial Pyramid Pooling)空间池化金字塔增加感受野,提升特征提取和融合能力,提高对电动车驾驶员是否佩戴头盔的检测能力。实验表明,在是否佩戴头盔检测任务中,改进后框架算法的全类别mAP(mean average precision)达到96.63%,比原框架提高2.4%;其中改进后佩戴头盔类别的AP(Average Precision)比原框架提高4%,未佩戴头盔类别AP比原框架提高1%;F(F-Measure)值比原算法均提高4%,改进后的算法更满足头盔佩戴检测任务。Aiming at the phenomenon of electric vehicle drivers not wearing helmets,an improved YOLOv4(You only look once)algorithm for detecting the helmet wearing of electric vehicle drivers is proposed.The K-means algorithm was used to cluster the data set to obtain a priori box,which improved the matching degree of the a priori box and the feature map.The convolutional layer was added to the output layer of the CSPDarknet53 backbone feature extraction network,and the SPP(Spatial Pyramid Pooling)space was added to the PANet(Path Aggregation Network)network part to increase the receptive field,improve the feature extraction and fusion capabilities,and improve the ability to detect whether the electric vehicle driver is wearing a helmet.Experiments show that in the task of detecting whether to wear a helmet,the full-category mAP(mean average precision)of the improved frame algorithm reaches 96.63%,which is 2.4%higher than the original frame;the improved AP(Average Precision)of the helmet type is 4%higher than the original frame,and the AP of the non-helmeted category It is 1%higher than the original framework;the F(F-Measure)value is 4%higher than the original algorithm.The improved algorithm is more suitable for helmet wearing detection tasks.

关 键 词：深度学习 头盔检测 特征提取 空间金字塔 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]