摘要
背景减除(background subtraction,BS)是森林火灾火焰图像中运动像素检测的主要手段,已有文献多采用高斯混合模型进行运动像素检测,然而GMM算法对于背景包含大量运动像素的视频并不能取得好的效果,而野外环境的森林火灾视频风吹树枝叶会导致大量的运动像素。为更好地提取森林火焰的运动像素,对36种BS检测算法应用于光照强度变化、树枝叶摆动、相机抖动等多种森林火灾视频进行实验,并选择4种效果好的BS检测算法(高斯平均值背景减除法(DPWren GA)、改进的高斯混合模型(DPZivkovic AGMM)、混合高斯背景模型(Mixture Of Gaussian V2)、局部二进制相似度分割背景减除法(LOBSTER))进行分析研究。结果表明,DPWren GA法检测到的前景完整性不够理想,但其噪声点最少,精确率最高。LOBSTER法能最完整地分割出前景。4种算法对阳光强烈的场景和存在树枝遮挡的火灾视频的检测效果都不是很理想,但对相机抖动有较好的适应性。
Forest is a valuable resource,but forest fire has a harmful effect on forest cover. In recent years,the methods of managing forest fire based on video images have been developed rapidly. Flame is an important visual feature of forest fire,so it is significant to regard the flame as a recognition standard of forest fire. Because of the flickering feature of the flame,detecting moving targets in the video is the basis of flame detection. Background subtraction( BS)is the main means of detecting moving pixels in forest fire flame images. Gaussian mixture model( GMM) has been used to detect moving pixels in the previous studies,but GMM algorithm cannot achieve good results from the videos whose background contains a lot of moving pixels. If forest fire videos with the wild scenes,wind blowing branches and leaves will lead to a large number of moving pixels. In order to extract the moving pixels of forest flame better,36 kinds of the BS algorithms were applied to experiments such as light intensity change,leaf wobble and camera shake,and 4 with good performances of which were selected for analysis. In this paper,the moving pixels of flame were extracted through the BS algorithms firstly,and then the flame area of forest fire was detected by examining the color feature of flame. Furthermore,precision rate,true positive rate and false positive rate were used as features to measure algorithm performance. The experimental results show that Gaussian average background subtraction( DPWrenGA)is not ideal enough for foreground integrity,but it has the least noise and the highest accuracy. Local binary similarity segmenter background subtraction( LOBSTER) can segment the foreground most completely. Four kinds of algorithms are not suitable for the scene of bright sunlight and the fire videos with branches covering,but they have better adaptability to the camera shake.
作者
陈越
赵亚琴
蒋林权
孙一超
CHEN Yue;ZHAO Yaqin;JIANG Linquan;SUN Yichao(Mechanical and Electronic Engineering College,Nanjing Forestry University,Nanjing 210037,China)
出处
《林业工程学报》
北大核心
2018年第4期131-136,共6页
Journal of Forestry Engineering
基金
国家自然科学基金青年基金(31200496)
关键词
森林火灾
火焰检测
运动检测
背景减除
forest fire
flame detection
motion detection
background subtraction