摘要
针对安全帽佩戴检测精度低等问题,提出一种基于YOLOv5的安全帽佩戴检测模型CA-YOLO.在YOLOv5骨干网络中引入坐标注意力机制,使模型可以注意到更多特征信息以提升检测精度,在训练过程中,提出一种基于正样本匹配与指数移动平均的PSM-EMA优化策略,并结合负样本训练降低模型误检率.在Helmet Detection和SHWD数据集上的实验结果表明,CA-YOLO的均值平均精度(x mAP)分别达到90.67%和91.02%,对佩戴安全帽工人的平均精度(y AP)达到94.53%和94.84%,相较于YOLOv5,该算法的均值平均精度分别提升2.42%和1.39%,CA-YOLO的平均精度均优于YOLOv5和其他对比模型,能实现不同复杂场景下对安全帽佩戴的准确检测.
This paper gives a CA-YOLO model of helmet wearing detection based on YOLOv5 for challenge of low detection accuracy on helmet wearing.The coordinate attention mechanism is introduced into YOLOv5 backbone network,so that the model can pay attention to more feature information to improve the detection accuracy.In the training process,a PSM-EMA optimization strategy based on positive sample matching and exponential moving average is proposed,and combined with negative sample training to reduce the module false detection rate.Experimental results based on datasets in Helmet Detection and SHWD show that the mean Average Precision(mAP)of CA-YOLO reaches 90.67%and 91.02%respectively,and the Average Precision(AP)for workers wearing helmets reaches 94.53%and 94.84%.Compared with YOLOv5,the mean average precision of the algorithm is increased by 2.42%and 1.39%respectively.The average accuracy of CA-YOLO is better than that of YOLOv5 and other contrast models,which can accurately detect helmet wearing in different complex scenes.
作者
崔海彬
蒲东兵
陆云凤
王敬
CUI Hai-bin;PU Dong-bing;LU Yun-feng;WANG Jing(School of Information Science and Technology,Northeast Normal University,Changchun 130117,China)
出处
《东北师大学报（自然科学版）》
CAS
北大核心
2023年第3期94-100,共7页
Journal of Northeast Normal University(Natural Science Edition)
基金
吉林省科技厅重点科技研发项目(20200401086GX)
长春市科学技术局资助项目(21ZY31).
关键词
安全帽
YOLO
坐标注意力机制
数据优化
目标检测
helmet
you only look once(YOLO)
coordinate attention mechanism
data enhancement
object detection
