【作者】 康锋；

【导师】 赵匀；

【作者基本信息】 浙江大学 ， 农业机械化工程， 2010， 博士

【摘要】 火灾是最普遍地威胁公众安全和社会发展的主要灾害之一,对农业以及工业和人们的生产、生活带来了极大的危害。随着科学技术的发展,火灾自动检测技术逐渐成为预防和监测火灾的有效手段。基于视觉特征的早期火灾自动探测技术,克服了传统火灾自动检测技术成本高、安装复杂,难以应用于农田、草场、森林等大范围火灾检测等缺点,可以更快捷、准确的检测火灾,环境适应能力强,而且结合计算机智能技术,可以提供更直观、更丰富的火灾信息。作为一种新型的、智能化的检测技术,基于视觉特征的火灾检测技术必将具有广阔的应用前景和使用价值。本文主要研究了基于视觉特征的早期农林火灾检测方法和相关算法。首先阐述了火灾检测系统的组成,分析了早期火灾图像的动态、静态特征,并对火灾图像的特征进行了算法流程分析；针对火灾图像的视觉特征和检测算法的实时性要求,提出了一种新的图像增强方法,扩大图像中不同物体之间的特征差别,同时保留了原图像的色彩信息,为火灾图像的信息提取及后续处理奠定基础；提出了一种基于神经网络的彩色图像分割算法,可以方便、准确的提取火焰和烟雾图像,为后续的火焰和烟雾像素点颜色判别规则提供统计信息；提出了一种基于GICA的快速运动检测算法,为后续的火焰和烟雾检测提供运动检测算法支持；提出了基于视觉特征的实时火焰检测方法和基于视觉特征的实时烟雾检测方法,对农林火灾检测进行了基础研究。本文的主要工作内容如下：1、介绍了火灾检测研究对预防农林火灾的重要意义,论述了开展基于视觉的农林火灾自动检测的重要性和必要性；分析了基于视觉特征的火灾检测技术与传统火灾检测的特点,并对两者的特点进行了比较,阐述了基于视觉特征的火灾检测技术优越性；分析了基于视觉特征的火灾自动检测系统组成,阐述了系统各部分的主要功能和作用。2、论述了基于视觉特征的火灾检测算法及研究现状；提出了基于视觉特征的火灾检测算法流程；分析了火焰图像的静态、动态特征,对目前火焰检测算法的研究现状进行了综述；分析了烟雾图像的特征,对目前烟雾检测算法的研究现状进行了综述。3、针对火灾图像的特征和检测算法实时性的要求,提出了一种基于模糊逻辑的彩色图像快速增强算法,保留了火灾图像的静态和动态特征,采用新的隶属度函数,对HSI彩色模式下的像素点进行处理,通过查表法完成了火灾图像的快速增强处理,提高了算法运算速度,以适应实时性要求。4、为了获取火灾图像像素点的颜色统计特征,提出了一种基于神经网络的彩色图像分割算法,以火灾图像中火焰或烟雾像素点的色彩和位置信息作为样本特征,采用多层前馈网络进行火焰图像的分割,提取了目标火灾图像。5、分析了目前在火灾检测中所采用的运动检测方法及其特点,提出了一种基于GICA的运动区域检测算法,克服了传统运动检测方法由于相邻帧灰度变化的带来的噪声,以及由于缓慢运动或运动目标与背景图像颜色相似而检测不到的缺点。6、根据火焰的视觉特征设计了实时火焰检测算法。对火焰图像序列进行GICA运动检测,提取运动像素点;建立了火焰色彩模型,滤除不具有火焰颜色特征的运动像素点；采用小波分析对运动像素点进行时域、频域分析,对多种火焰时频特征建立了判别规则。7、根据烟雾的视觉特征的设计了实时烟雾检测算法。对烟雾图像序列进行GICA运动区域检测,提取运动图像；根据烟雾运动图像的颜色分布特征,建立了烟雾像素点判别规则；采用小波对运动像素点进行时域、频域分析,对多种烟雾时频特征建立了判别规则。8、最后总结了本文的研究内容,指出了本文的创新点,提出了进一步的研究方向。更多还原

【Abstract】 Fire is the one of the universal threats to public safety and social development, and brings great harm to productivity in agriculture, industry and people’s life. With the development of science and technology, automatic fire detection has become an effective means of fire prevention and monitoring. Vision-based early fire automatical detection technology can be more efficient and accurate to detect fires with great environmental adaptation ability. Combined with computer technology, it can provide more intuitive and abundant fire information. Compared with the vision-based fire automatical detection technol ogy, the tradi ti onal automati c fi re detecti on technology is costly and complicated to install, difficult to apply to farmland, pasture, forest fire detection and other situations of large space. As a new and intelligent method, fire detection technique based on vision will have broad application value and prospects.This thesis studies the early fire detection method based on visual characteristics of the fire. First, it describes the composition of the fire detection system, analyses the static and dynamic characteristics of the wild early fires, and analyses the procedure of the fire detection algorithm. I n order to meet the real-time requirements of the detection algorithm, thi s thesi s proposes a new i mage enhancement method to enhance the characteri sti cs of the image between the different objects for image information extraction subsequently. This thesis also proposes a method of the color image segmentation based neural network, extracts the pixels of flames and smoke images, and provides a statistical rule to discriminate flames and smoke for the subsequent analysis of pixel color. This thesis proposes a new algorithm of fast motion detection to support the subsequent processing of flame and smoke detection. At last, this thesis proposes the methods of real-time flame detection and real-time smoke detection based on visual characteristics, and completes the basic research of fire detection.The main work of this thesis as follows:1. This thesis introduces the meaning of fire detection and prevention in the agricultural. First, this thesis discourses on the importance and necessity of the automatic fire detection, compares the features of fire detection technology based on visual characteristics with the traditional fire detection technology, and then expounds the advantages of the fire technology based on visual features. Finally, this thesis analyzes the components of the automatic fire detection system based on visual features, and explained the main f uncti ons and the role of the vari ous parts of the system.2. This thesis discusses and compares the fire detection algorithms based on visual features and current situations correspondingly, proposes the algorithm processes of the fire detection based on visual features, analyses the static and dynamic visual characteri stics of the smoke and flame image, and reviews the detection algorithm of flame and smoke detection methods.3. This thesis proposes a fast algorithm of color image enhancement based on fuzzy logic algorithm in order to meat the real-time demands for the practical application. This algorithm preserves the static and dynamic characteristics of the fire image, and it processes to the pixel directly in the HSI color model with the new membership function, complete the rapid enhancement processing through the lookup table to the fire image to meet the real-time requirements.4.In order to obtai n the statistical characteri sti cs of the pixel s of f i re i mage, thi s thesi s proposed an algorithm of color image segmentation based a neural network. In this algorithm, the color and location information of fire pixels are chosen as sample characteristics, multilayer feed forward neural networks are applied for image segmentation to extract the target of f i re i mages.5. This thesis analyses the current methods of moving detection in fire detecti on, and proposed a moving detection algorithm based on GICA. This algorithm overcomes the shortcomings of the traditional method of the moving detection that bring about noise because of changes i n the adjacent frame gray and inefficient to slow movement.6. This thesis proposes a method of real-time flame detecti on based on visual features of the flame image. First, GICA algorithm is applied to the flame image sequences to extract motion pixels, and then the flame color model establishes and filters the moving pixels movement without the characteristics of flame color pixel. Finally, some classification rules of flame detection establish based on time-frequency domain analysis by wavelet for the f i re al arm.7. This thesis proposes a method of real-time smoke detecti on based on visual features of the smoke image. GICA is applied to the smoke image sequence to extract the moving pixels, and the classificati on rul es establish according to the color of the smoke distribution. Finally, some classification rules of smoke detection establish based on the time-frequency characteristics of the motion pixels by wavelet for the fire alarm.8. At last this thesis summarizes the contents and the innovation of this study, and proposes the further research.更多还原