摘要
为解决传统传感器在检测火灾的过程中受到环境、安装距离等因素影响导致适应性差的缺点,本文基于视觉传感器,通过视觉目标检测技术对火灾进行检测,从而实现火灾的预警。首先,为了提高轻量级的目标检测网络(You Only Look Once v4 Tiny,YOLOv4-Tiny)在检测火灾目标时的准确率,本文基于森林火灾的数据集,运用二分K-Means聚类算法重新生成锚定框(Anchor Box)。然后,在传统YOLOv4-Tiny网络的基础上通过增加大尺度预测结果的方式,降低漏检率。最后,本文结合预训练权重重新训练火灾检测网络,并在英伟达板卡上进行部署实验。实验结果表明,本文的火灾检测网络在测试数据集上的准确率为97.81%,漏检率为4.83%,与原始YOLOv4-Tiny相比,准确率提高了3.13%,漏检率降低了6.44%,检测速度达到了16帧/s,综合性能良好,满足火灾检测的需求。
Aiming at the disadvantage of poor adaptability of traditional sensors due to factors such as environment and installation distance in the process of detecting fires,this paper detected the fire through visual target detection technology based on visual sensor,so as to realize fire warning.Firstly,the binary K-Means clustering algorithm was used to regenerate the Anchor Box to improve the detection accuracy of the lightweight target detection network(You Only Look Once v4 Tiny,YOLOv4-Tiny)according to dataset in this paper.Then,the large-scale prediction results were added on the basis of the traditional YOLOv4-Tiny network to reduce the missed detection rate.Finally,combined with the pre-training weights,fire detection network was retrained and deployed on NVIDIA board.The experimental results showed that the accuracy rate of the fire detection network in this paper was 97.81%and the missed detection rate was 4.83%on the test data set.The accuracy rate had increased by 3.13%and the missed detection rate had been reduced by 6.44%compared with the original YOLOv4-Tiny.The detection speed had reached 16 frames per second.The overall performance was good and met the needs of fire detection.
作者
缪伟志
陆兆纳
王俊龙
王焱
MIAO Weizhi;LU Zhaona;WANG Junlong;WANG Yan(School of Automotive Engineering,Nantong Institute of Technology,Nantong 226002,China)
出处
《森林工程》
北大核心
2022年第1期86-92,100,共8页
Forest Engineering
基金
2021年南通市基础科学研究项目(JC2021065)。
