摘要
传统视频火灾识别以人工特征为依据和以图像处理技术为主,针对其受复杂背景干扰较大导致误报率较高等缺点,提出一种基于深度学习的火灾识别算法.首先采用混合高斯建模法进行运动检测,然后采用集成学习中Adaboost算法对运动图像进行疑似火灾区域提取,最后采用轻量级神经网络MobileNetv3自动提取疑似火灾区域特征进行火灾识别.仿真结果表明,本算法的平均准确率可达98.1%,平均误报率为4.9%,能适应不同的复杂场景,并且保持较好的实时性.
Aiming at the vulnerability to interference of complex background and high false alarm rate of the traditional video fire detection,which is based on artificial features and the image processing technology,a fire detection algorithm based on deep learning was proposed.First,the Gaussian mixture modeling method was used for motion detection.Then,the Adaboost algorithm was used to extract the suspected fire area from the moving image.Finally,the lightweight neural network MobileNetv3 was used to automatically extract the suspected fire area feature for fire detection.The simulation results show that the average accuracy of this algorithm can reach 98.1%,with an average low false alarm rate of 4.9%.Besides,the algorithm can be applied to different complex scenarios and maintain a good real-time performance.
作者
陈培豪
肖铎
刘泓
Chen Peihao;Xiao Duo;Liu Hong(Zhejiang University City College,Hangzhou 310015,China;College of Control Science and Engineering,Zhejiang University,Hangzhou 310027,China)
出处
《燃烧科学与技术》
CAS
CSCD
北大核心
2021年第6期695-700,共6页
Journal of Combustion Science and Technology
基金
浙江省重点研发计划资助项目(2019C01150).
关键词
深度学习
特征提取
火灾识别
deep learning
feature extraction
fire detection