基于烟雾运动特征的高速公路火灾检测研究  

Research on Highway Fire Detection Based on Smoke Motion Characteristics

作　　者：李许峰 成高立 梁浩翔 宋焕生[3] LI Xufeng;CHENG Gaoli;LIANG Haoxiang;SONG Huansheng(Shaanxi Expressway Mechanisation Engineering Co.,Ltd.,Xi’an 710038,China;School of Electronics and Control Engineering,Chang’an University,Xi’an 710064,China;School of Information Engineering,Chang’an University,Xi’an 710064,China)

机构地区：[1]陕西高速机械化工程有限公司,陕西西安710038 [2]长安大学电子与控制工程学院,陕西西安710064 [3]长安大学信息工程学院,陕西西安710064

出　　处：《微型电脑应用》2024年第6期5-8,共4页Microcomputer Applications

基　　金：国家自然科学基金(6207072223)。

摘　　要：目前高速公路火灾检测方法仍存在检测精度低、计算资源消耗多的问题。利用烟雾的运动特征结合深度网络,设计一个高速公路火灾检测方法。利用高斯混合背景建模提取视频中的运动前景,对运动前景进行感兴趣区域筛选,得到疑似烟雾区域,再对该区域进行HSV颜色空间分析,确定该区域是否存在烟雾。对于存在烟雾的视频帧,通过构建的具有5万余个目标标注框的高速公路专用火灾烟雾数据集,利用YOLOv5目标检测方法,达到了90.16%的平均检测精度(mAP)。所提方法避免了逐帧对视频流的火灾检测,极大地减少了计算资源的浪费,具有研究与实际工程应用价值。The current fire detection methods still have the problems of low detection accuracy and high computational resource consumption.This paper designs a highway fire detection method based on the motion of smoke combined with deep network.The moving foreground in the video is extracted by Gaussian mixture background modeling,and the region of interest is selected for the moving foreground to obtain the potential smoke area,and the HSV color space analysis is performed on the area to determine the presence of smoke.For video frames with smoke,a highway-specific fire and smoke dataset with more than 50000 object annotation boxes is constructed,combined with the YOLOv5 detection method,to achieve a mean average precision(mAP)of 90.16%.The proposed method avoids frame-by-frame fire detection,greatly reduces the waste of computing resources,and has research and practical engineering application values.

关 键 词：烟雾检测 火灾烟雾数据集 交通视频分析 深度学习 高速公路火灾 

分 类 号：P315.69[天文地球—地震学]