中国安全科学学报 ›› 2023, Vol. 33 ›› Issue (6): 152-158.doi: 10.16265/j.cnki.issn1003-3033.2023.06.1532

• 公共安全 • 上一篇    下一篇

基于YOLOv5s的城镇森林交界域火灾探测模型

王喆1,2(), 李享1,2, 杨栋梁1,2,**(), 刘丹1,2   

  1. 1 武汉理工大学 安全科学与应急管理学院,湖北 武汉 430070
    2 武汉理工大学 中国应急管理研究中心,湖北 武汉 430070
  • 收稿日期:2023-01-15 修回日期:2023-04-09 出版日期:2023-08-07
  • 通讯作者:
    **杨栋梁(1996—),男,河南驻马店人,硕士研究生,研究方向为应急情报分析。E-mail:
  • 作者简介:

    王喆 (1980—),男,湖北武汉人,工学博士,副教授,博士生导师,主要从事应急决策、人工智能、应急物流等方面的研究。E-mail:

  • 基金资助:
    教育部人文社会科学研究青年基金资助(20YJC630154); 国家自然科学基金资助(62073251); 国家自然科学基金资助(71501151); 湖北省自然科学基金资助(2020CFA055); 湖北省自然科学基金资助(2016CFB467); 中央高校基本科研业务费专项资金资助(2020-VI-004)

Fire detection model of wildland-urban interface based on YOLOv5s

WANG Zhe1,2(), LI Xiang1,2, YANG Dongliang1,2,**(), LIU Dan1,2   

  1. 1 School of Safety Science and Emergency Management, Wuhan University of Technology, Wuhan Hubei 430070, China
    2 China Emergency Management Research Center, Wuhan University of Technology, Wuhan Hubei 430070, China
  • Received:2023-01-15 Revised:2023-04-09 Published:2023-08-07

摘要:

为精准监控城镇森林交界域火灾及定位其空间分布,提出基于改进YOLOv5s网络的城镇森林交界域火灾目标探测模型。首先收集城镇森林交界域火灾图像,利用图像注释工具标注出目标探测数据集;然后将坐标注意力(CA)机制引入YOLOv5s的主干网络,增强模型的方向及位置信息感知,以精准定位出城镇森林交界域火灾起火点;最后以准确度、召回率、平均准确度为评价指标,在自建数据集上进行训练、测试。模拟结果表明:改进的YOLOv5s模型整体性能提升,在城镇森林交界域火灾目标探测中,建筑物火灾平均精确度增加了0.8%,森林火灾则增加了1.3%。

关键词: 城镇森林交界域, YOLOv5s, 火灾探测, 目标探测

Abstract:

In order to accurately monitor the fires at wildland-urban interfaces and locate their spatial distribution, a target detection model of wildland-urban interface fires based on the improved YOLOv5s network was proposed. Images of fires at the wildland-urban interfaces were collected, and object detection datasets were annotated with the image annotation tool. CA mechanism was introduced into the backbone network of YOLOv5s to enhance the orientation and location information perception of the model to accurately locate the fire point at the wildland-urban interface. Based on the evaluation indicators of accuracy, recall rate and average accuracy, training and testing were carried out on the self-built data set. The experimental results show that the overall performance of the improved YOLOv5s model is improved, and the average accuracy of building fires increases by 0.8% and forest fires by 1.3% in the detection of fire targets in the wildland-urban interface.

Key words: wildland-urban interface, YOLOv5s, fire detection, target detection