中国安全科学学报 ›› 2023, Vol. 33 ›› Issue (2): 31-37.doi: 10.16265/j.cnki.issn1003-3033.2023.02.0298

• 安全社会科学与安全管理 • 上一篇    下一篇

基于YOLO-V3算法的加油站不安全行为检测

常捷1,2(), 张国维1,2, 陈文江3,4, 袁狄平1, 王永生5   

  1. 1 中国矿业大学 深圳研究院,广东 深圳 518057
    2 中国矿业大学 安全工程学院,江苏 徐州 221116
    3 深圳市城市公共安全技术研究院有限公司 科信中心,广东 深圳 518000
    4 中国科学院 深圳先进技术研究院,广东 深圳 518055
    5 江苏鸿鹄安全科技有限公司 研发中心,江苏 徐州 221116
  • 收稿日期:2022-09-19 修回日期:2022-12-10 出版日期:2023-02-28 发布日期:2023-08-28
  • 作者简介:

    常捷(1996—),男,河南郑州人,硕士研究生,主要研究方向为智慧消防、城市新型火灾防护理论与技术。E-mail:

    张国维,副教授

    袁狄平,教授

  • 基金资助:
    山东省重点研发计划(2021CXGC011303); 广东省重点领域研发计划项目(2019B111102002); 江苏省自然科学基金面上项目资助(BK20221548)

Gas station unsafe behavior detection based on YOLO-V3 algorithm

CHANG Jie1,2(), ZHANG Guowei1,2, CHEN Wenjiang3,4, YUAN Diping1, WANG Yongsheng5   

  1. 1 Shenzhen Research Institute, China University of Mining and Technology, Shenzhen Guangdong 518057, China
    2 School of Safety Engineering, China University of Mining and Technology, Xuzhou Jiangsu 221116, China
    3 Science and Technology Information Center, Shenzhen Urban Public Safety and Technology Institute, Shenzhen Guangdong 518000, China
    4 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen Guangdong 518055, China
    5 Research and Development Center, Jiangsu Honghu Safety Technology Co., Ltd., Xuzhou Jiangsu 221116, China
  • Received:2022-09-19 Revised:2022-12-10 Online:2023-02-28 Published:2023-08-28

摘要:

为控制加油站火灾爆炸风险目标,结合事故统计和故障树分析方法,提出一种基于YOLO-V3算法的加油站不安全行为检测模型。首先在收集90起加油站火灾爆炸事故的基础上,统计分析加油站火灾爆炸事故的点火源;其次构建加油站火灾爆炸故障树,计算各基本事件的结构重要度,并确定加油站危险性较高的不安全行为;然后采用现场采集和模拟的方法收集加油站不安全行为图像数据,利用数据增强方法构建加油站不安全行为图像数据集;最后基于深度学习的方法构建加油站不安全行为检测模型,经过1000次训练迭代后得到最终模型。研究结果表明:引起加油站火灾爆炸事故的不安全行为主要有抽烟、打电话等;训练得到的检测模型在测试集上对抽烟、打电话和正常行为检测类别的平均检测精度分别为67%、85%和77%,模型的平均检测精度均值为84%。

关键词: YOLO-V3算法, 加油站, 故障树, 不安全行为, 火灾爆炸, 目标检测

Abstract:

In order to control the fire and explosion risk target of gas station, an unsafe behavior detection model of gas station based on YOLO-V3 algorithm was proposed by combining accident statistics and fault tree analysis,. Firstly, on the basis of collecting 90 gas station fire and explosion accidents, the ignition sources of gas station fire and explosion accidents were statistically analyzed. Secondly,. the fire and explosion fault tree of gas station was constructed, the structural importance of each basic event was calculated, and the unsafe behavior with high risk of gas station was determined. Then, the image data of unsafe behavior of gas stations were collected by field collection and simulation, and the image data set of unsafe behavior of gas stations was constructed by data enhancement method. Finally, based on the deep learning method, the unsafe behavior detection model of gas station was constructed, and the final model was obtained after 1000 training iterations. The results show that the unsafe behaviors that cause fire and explosion accidents in gas stations mainly include smoking, calling and so on. The average detection accuracy of the trained detection model for smoking, calling and normal behavior detection categories on the test set is 67%, 85% and 77%, respectively, and the average detection accuracy of the model is 84%.

Key words: YOLO-V3 algorithm, gas station, fault tree, unsafe behavior, fire and explosion, target detection