【作者】 郭炜强；

【导师】 韩宁； 燕飞；

【作者基本信息】 北京林业大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 森林是人类生存与发展不可或缺的资源,林火是造成森林资源破坏的最主要灾害。因此及时发现火情,对尽早扑灭林火减少森林资源损失有着重要的意义。目前,我国大部分林区虽然都建立了视频林火监控网络,但火情判断主要还是依靠人工观测监控。在现有的硬件系统基础上加入林火智能识别算法,不仅能够实现火情自动报警,而且改造成本较低。烟雾是森林火灾前期最为明显的特征,所以基于可见光视频的烟雾智能识别研究具有更加重要的意义。本文通过理论仿真分析与工程试验,提出了一个森林火灾烟雾识别算法框架。首先根据烟雾的运动特性,利用运动检测算法确定烟雾疑似区域,再对疑似区域提取图像特征,使用训练好的分类器进行智能识别排除干扰物影响。主要成果如下：(1)提出灰度投影算法对林火视频序列图像进行电子稳像,减小了大风造成的摄像机抖动影响。(2)引入运动检测法中的高斯模型背景减除法,利用火灾烟雾持续缓慢的扩散运动特性进行烟雾疑似区域提取,再使用带阈值的最小距离聚类法对零散琐碎区域聚合整理。(3)基于烟雾较强的视觉特征,提取10维烟雾图像特征,利用Adaboost算法通过对训练样本集合的不同组合训练出多个弱分类器——分类与回归决策树。最终通过对弱分类器的组合达到强分类器的效果,解决了烟雾识别非线性可分问题。仿真实验达到漏报率4.53%、误报率5.91%。(4)设计开发了森林火灾智能识别软件系统。在视频监控平台上融合了基于本文研究的序列图像去抖动、运动目标检测、图像特征提取以及Adaboost算法分类识别方法等,开发了森林火灾智能识别系统。经工程应用检验,系统运行稳定,识别准确率99.32%、误报率0.67%并成功捕获一次真实火情。更多还原

【Abstract】 Forest is an important resource which is essential to human being’s survival and development. As a major nature disaster, the forest fire damages large quantity of forest resources very year. In order to put out the forest fire as fast as possible and reduce the loss, it makes sense to detect the fire in time. Nowadays, video network has been established in most forests to monitor fire. However, it purely relies on artificial recognition. Forest fire video monitoring network based on hardware and advanced detection algorithm can not only automatically alarm of fire, but also decrease the cost significantly. Smoke is the most significant phenomenon prior to a fire, so research on video-based smoke recognition makes a good sense.In this paper, the author proposed a frame of forest fire smoke recognition system. Several major results of this study are as follows:(1) Gray projection algorithm which is one of electronic image stabilization methods is put forward to maintain the video series smooth and stable. The algorithm has good effect on image sequence with colorful and small timing jitter.(2) Moving target detection algorithm based on Gaussian statistical model is proposed. Scattered trivial regions are clustered by minimum distance clustering with threshold.(3)10dimension image features of forest fire smoke are extracted based on distinguishing visual characteristic of smoke. Multiple weak classifiers, classification and regression tree, is constructed with Adaboost algorithm. Strong classifier is established combined with those weak classifiers, and could be used to detect forest fire smoke. Simulation results show that the identification ratio is above94.7%, while false positive ratio was4.53%, false negative ratio was5.91%.(4) Integration of wireless networks, video monitoring and other techniques contributes to the intelligent forest fire smoke recognition system. This system implements the results of this study, including electronic image stabilization, forest fire video motion detection algorithm, image feature extraction method and Adaboost classification and recognition method. It was deployed in a forest in Beijing, and successfully captured a true forest fire and all simulated forest fires tests. The identification ratio was99.32%.更多还原