计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (11): 213-220.DOI: 10.3778/j.issn.1002-8331.1811-0389

• 工程与应用 • 上一篇    下一篇

改进YOLO v3的安全帽佩戴检测方法

施  辉,陈先桥,杨  英   

  1. 武汉理工大学 计算机科学与技术学院,武汉 430063
  • 出版日期:2019-06-01 发布日期:2019-05-30

Safety Helmet Wearing Detection Method of Improved YOLO v3

SHI Hui, CHEN Xianqiao, YANG Ying   

  1. College of Computer Science and Technology, Wuhan University of Technology, Wuhan 430063, China
  • Online:2019-06-01 Published:2019-05-30

摘要: 在生产和作业场地中,工人由于不佩戴安全帽而引发的安全事故时有发生。为了降低由于未佩戴安全帽而引发的安全事故发生率,提出了一种基于改进YOLO v3算法的安全帽佩戴检测方法。通过采用图像金字塔结构获取不同尺度的特征图,用于位置和类别预测;使用施工现场出入口监控视频作为数据集进行目标框维度聚类,确定目标框参数;在训练迭代过程中改变输入图像的尺寸,增加模型对尺度的适应性。理论分析和实验结果表明,在安全帽佩戴检测任务中,mAP(Mean Average Precision)达到了92.13%,检测速率提高到62?f/s,其检测准确率与检测速率相较于YOLO v3均略有提高,所提算法不仅满足安全帽佩戴检测中检测任务的实时性,同时具有较高的检测准确率。

关键词: 图像处理, 深度学习, YOLO v3, 安全帽佩戴检测

Abstract: In production and operation sites, safety accidents caused by workers not wearing safety helmets occur from time to time. In order to reduce such incidence, a safety helmet wearing detection method based on improved YOLO v3 algorithm is proposed. Firstly, feature maps of different scales are acquired by using an image pyramid structure for position and category prediction. Secondly, the target frame parameters are clustered by using the construction site entrances and exits monitoring video as the data set. Finally, during the training process, the size of the input image is changed during the iterative process to increase the adaptability of the model to the scale. Theoretical analysis and experimental results show that in the safety helmet wearing detection task, the mAP(Mean Average Precision) reaches 92.13%, and the detection rate increases to 62 frame/s. The detection accuracy and detection rate are slightly improved compared with YOLO v3. It not only meets the real-time performance of the detection task in the safety helmet wearing test, but also has a high detection accuracy.

Key words: image processing, deep learning, YOLO v3, safety helmet wearing detection