文献类型：期刊文献

中文题名：基于视频区域动态特征的林火烟雾检测技术研究

英文题名：Research on forest fire smoke detection technology based on video region dynamic features

作者：刘长春[1,2] 刘鹏举[1,2] 季烨云[1,2]

第一作者：刘长春

机构：[1]中国林业科学研究院资源信息研究所,北京100091;[2]国家林业和草原局林业遥感与信息技术重点实验室,北京100091

年份：2021

卷号：43

期号：1

起止页码：10-19

中文期刊名：北京林业大学学报

外文期刊名：Journal of Beijing Forestry University

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；

基金：中国林业科学研究院基本科研业务费专项(CAFYBB2017ZC001);“948”国家林业局引进项目(2014-4-01)。

语种：中文

中文关键词：林火烟雾检测;视频图像;分区;信噪比;LBP纹理特征

外文关键词：forest fire smoke detection;video image;block;signal-to-noise ratio;LBP texture feature

分类号：S762.3+2

摘要：【目的】视频监控越来越多地应用到森林火灾烟雾的早期检测中。现有的视频林火烟雾检测方法大多是基于像素提取烟雾特征进行分析检测,烟雾发生早期或烟雾距离摄像头较远时,在视频图像上烟雾仅呈现较小区域,且烟雾的扩散具有无规则性,背景环境复杂多变,导致基于像素的特征不明显,因此使基于像素的烟雾自动化检测难度增大。本文根据可见光视频图像处理原理,提出一种基于局部区域图像动态特征的林火视频烟雾检测方法,以提高林火视频烟雾检测准确度和灵敏度。【方法】以视频图像为研究对象,每秒取一帧生成图像序列,对图像序列进行多层次不同尺度分区;利用图像信噪比原理,计算分区后的连续图像序列的信噪比;根据背景图像信噪比得到自适应阈值,确定待检测图像序列发生亮度变化的图像块,即为疑似烟雾块;提取疑似烟雾块的LBP纹理特征,采用支持向量机区分出烟雾区域。【结果】利用HSV颜色空间的亮度分量,可以有效提取烟雾区域。选择有林火烟雾的视频,对提出的烟雾变化检测方法进行验证,分析结果表明该方法能确定烟雾发生所在的图像块,且能排除部分非烟雾干扰因素。【结论】本文提出了基于局部区域亮度特征和LBP纹理特征的视频林火烟雾检测技术,能准确定位烟雾发生区域,排除部分干扰因素,检测识别率平均达到92%以上,有助于实时林火烟雾自动检测,提高林火烟雾检测率,具有很强的实用性。
[Objective]Video surveillance is increasingly applied to the early detection of forest fire smoke.The existing video forest fire smoke detection methods are mostly based on pixel extraction of smoke characteristics for analysis and detection,but when the smoke is early or the smoke is far from the camera,the smoke only appears in a small area on the video image.Moreover,the diffusion of smoke is irregular,and the background environment is complex and changeable,resulting in insignificant pixel-based features,which makes it more difficult to automatically detect pixel-based smoke.Based on the principle of visible light video image processing,this paper proposes a forest fire video smoke detection method based on local area image dynamic characteristics to improve the accuracy and sensitivity of forest fire video smoke detection. [Method] The video images were selected as the research object. One frame per second wastaken to generate an image sequence, and the image sequences were divided into multiple levels anddifferent scales;using the principle of image signal-to-noise ratio, we calculated the signal-to-noise ratio ofcontinuous image sequences after blocking;the adaptive threshold was obtained according to the signal-tonoiseratio of the background image, and the image block whose brightness changes in the image sequenceto be detected was determined to be the suspected smoke block;the LBP texture feature of the suspectedsmoke block was extracted, and the support vector machine was used to distinguish the smoke area.[Result] Using the value component of the HSV color space, smoke areas can be effectively extracted. Thevideos with forest fire smoke were selected to verify the proposed smoke change detection method. Theanalysis results showed that the method can determine the image block where the smoke occurred andexcluded some non-smoke interference factors. [Conclusion] This paper proposes a video forest fire smokedetection technology based on brightness characteristics and LBP texture features of local area, which canaccurately locate the smoke occurrence area and exclude some interference factors. The average detectionrecognition rate reaches more than 92%, which is helpful for real-time forest fire smoke automatic detectionand improving the detection rate of forest fire smoke. It has a strong practicality.