英文版论文原文：https://ieeexplore.ieee.org/document/1421401?arnumber=1421401
一种基于图像处理的早期火灾探测方法
An early fire-detection method based on image processing
Thou-Ho Chen, Ping-Hsueh Wu& Yung-Chuen Chiou
Abstract
本文提出了一种基于视频处理的早期火警预警方法。提出火灾探测的基本思想是采用基于RGB（红色，绿色，蓝色）模型的色度和无序测量来提取火灾像素和烟雾像素。火象素的决策函数主要由R分量的强度和饱和度推导。通过生长和无序动态以及进一步的烟雾，可以验证提取的火像素是否为真火。基于对火焰增长比率的迭代检查，在满足警报条件时发出火警。实验结果表明，所开发的技术能够以较低的误报率实现火灾事故的全自动监控，因此在重要的军事，社会保障，商业应用等方面具有很大的吸引力，
The paper presents an early fire-alarm raising method based on video processing. The basic idea of the proposed of fire-detection is to adopt a RGB (red, green, blue) model based chromatic and disorder measurement for extracting fire-pixels and smoke-pixels. The decision function of fire-pixels is mainly deduced by the intensity and saturation of R component. The extracted fire-pixels will be verified if it is a real fire by both dynamics of growth and disorder, and further smoke. Based on iterative checking on the growing ratio of flames, a fire-alarm is given when the alarm-raising condition is met. Experimental results show that the developed technique can achieve fully automatic surveillance of fire accident with a lower false alarm rate and thus is very attractive for the important military, social security, commercial applications, and so on, at a general cost.
INTRODUCTION
通常，火灾事故经常造成经济和生态破坏，并危及人们的生命。为了避免火灾的发生，人们探索了许多早期的火灾探测技术，除了传统的紫外线和红外火灾探测器以外，大多数技术还基于粒子采样，温度采样，相对湿度采样，空气透明度测试，烟雾分析。但是，这些探测器要么必须放在火源附近，要么不能提供有关燃烧过程的其他信息，例如火源的位置，大小，生长速率等。因此，它们并不总是可靠的，因为可能会因故障而检测到以其他方式产生的非火或燃烧副产物的能量排放。这通常会导致错误警报。
Generally, the fire accident frequently causes economical and ecological damage as well as endangering people’s lives. To avoid the fire’s disasters, many early fire-detection techniques have been explored and most of them are based on particle sampling, temperature sampling, relative humidity sampling, air transparency testing, smoke analysis, in addition to the traditional ultraviolet and infrared fire detectors. However, those detectors either must be set in the proximity of a fire or can’t provide the additional information about the process of burning, such as fire location, size, growing rate, and so on. Thus, they are not always reliable because energy emission of non-fires or byproducts of combustion, which can be yielded in other ways, may be detected by misadventure. This usually results in false alarms. To provide more reliable information about fires, the visual-based approach is becoming more and more interesting.
就火灾特征的分割而言，与灰度处理相比，色彩处理由于光照条件（例如自然背景照明）的变化而产生的虚假警报更少。为了提高夜间的火警探测能力，Cappellini等人。[1]介绍了彩色视频以识别烟雾中的火焰。最近基于彩色视频的研究[2] – [3] [4] [5]提出了一些增强的彩色图像处理技术，以实现对火焰的实时检测。但是，以上方法都着重于对火灾的识别，但是无法提供对真实火灾的可靠验证，也无法提供有关火焰会燃烧还是燃烧的任何信息。考虑到商业成本，这非常重要，因为操作人员必须手动验证每个错误警报。为了降低森林火灾探测中的误报率，提出了一个复杂的混合系统，该系统具有多个由视觉摄像机，红外摄像机，气象传感器和地理信息数据库提供的输入[6]。。在不失一般性的前提下，混合方法总是带来更高的成本和组合维护。这激发了有吸引力的火灾探测方法可以针对通用，高可靠性和低成本特征[7]。它使用两阶段决策策略，第一个决策阶段将检测是否存在火灾，第二个决策阶段将进一步检查火灾是否会在以后蔓延。无论如何，某些火灾别名可能会使第一个决策阶段失败，并且需要完善第二个决策阶段以降低错误的火警发生率。
In point of the segmentation of fire features, the color processing will have less false alarms from the variation of lighting conditions, e.g., natural background illumination, than that of the gray-scale processing. To increase the fire-detection capability during the night, Cappellini et al. [1] introduces the color video to recognize the fire flame from smokes. Recent color-video based researches [2]–[3][4] [5], propose some enhanced color image processing techniques for achieving a real-time detection of fire flame. However, the above methods all focus on recognition of a fire but can’t provide a reliable validation of a real fire and any information about whether the flame will burn up or low. This is very important when the commercial cost is considered, since human operators must manually validate each false alarm. To reduce false alarm rate in forest-fire detection, a complex hybrid system with multiple inputs provided by the visual camera, the infrared camera, meteorological sensors and a geographical information database is presented [6]. Without losing the generality, a hybrid approach always brings a higher cost and maintenance on combination. This motivates that the attractive fire-detection method may be aimed at the general purpose, high reliable and low cost features [7]. It uses a 2-stage decision strategy that the first decision stage is to detect if there is a existing fire and the second decision stage is to further check whether the fire will spread out or not later on. Anyway, some fire aliases may make the first decision stage to be failed and the second decision stage needs to be refined for reducing false fire-alarm rate.
为了克服上述先前的问题，我们通过通过RGB色彩分割和乱序测量来验证提取的火灾像素和烟雾像素，从而改进了真实火灾的验证。火象素的判定函数是通过推导R分量的强度和饱和度而得到的，R，G，B相互比较，并利用强度推论出烟雾的强度。基于迭代的高可靠性火焰检查，可以实现火焰是否会燃烧或烧毁。如果认为有火在燃烧，将立即发出火警，以免引起火灾。
To overcome the previous problems mentioned above, we improve a real fire validation by verifying the extracted fire-pixels and smoke-pixels through RGB chromatic segmentation and disorder measurement. The decision function of fire-pixels is obtained by deducing with the intensity and saturation of R component, and the R, G, B as compared with each other and intensity is utilized to deduce that of the smoke. Based on iterative high reliable checking of flame if a fire flame will burn up or down can be achieved. If a fire is considered to be burning up, a fire alarm will be immediately given became the fire may lead to a disaster.
扫描二维码关注公众号，回复： 8572443 查看本文章
FIRE FEATURES
大多数燃料将在适当的条件下燃烧，会与空气中的氧气发生反应，生成燃烧产物，发光并释放热量。火焰是一种气相现象，很显然，液体和固体燃料的燃烧燃烧必须包括将其转化为气体形式。在大火方面[8]，火焰通常显示红色。此外，火焰的颜色会随着温度的升高而变化。着火温度低时，颜色显示范围从红色到黄色，而温度较高时，颜色可能变为白色。这表明，低温火焰发出高色彩饱和度的光，而高温火焰发出低饱和度的光。此外，白天或使用额外的光源时，火的色度比夜间或不使用光源时的火色具有更强的饱和度。应该指出的是，温度很高的火焰和某些特殊的可燃材料都可能产生蓝色火焰。火灾的另一个特征是由于风引起的气流会使火焰振荡或突然移动，因此形状发生了变化，如图所示。图一。基于视频处理，此动态功能将反映相应的效果，尤其是在图像中的可变火焰区域上。此外，由于燃烧不同的可燃燃料，烟总是在燃烧的火焰中产生，并具有不同的数量和颜色。基于以上对火灾的分析，这些功能将用于识别真实火灾。
Most fuels will burn under appropriate conditions, reacting will oxygen from the air, generating combustion products, emitting light and releasing heat. Flame is a gas phase phenomenon and, clearly, flaming combustion of liquid and solid fuels must involve their conversion to gaseous form. In the point of general fires [8], the flames usually display reddish colors; besides, the color of the flame will change with the increasing temperature. When the fire temperature is low, the color shows range from red to yellow, and it may become white when there is a higher temperature. This reveals that a low-temperature flame emits a light of high color’s saturation and a high-temperature flame emits a low-saturation light. Furthermore, the color of fires during the day or with the extra light source has a stronger saturation than that of during the night or no light source. It should be pointed out that both the flame with a very high temperature and some special combustible materials may generate bluish flame. Another feature of fires demonstrates the changeable shapes due to the fact that airflow caused by wind will make flames oscillate or move suddenly as shown in Figure I. Based on video processing, this dynamic feature will reflect the corresponding effect especially on a variable flame area in an image. Besides, smokes are always generated with a burning fire and have various quantities and colors because of burning different combustible fuels. Based on the above analyses of fire, these features will be used to identify a real fire.
2.1火焰色度分析
2.1 Chromatic Analysis of Flames
为了模拟人类视觉系统的色彩感应特性，通常将RGB颜色信息转换为数学空间，该数学空间将亮度（或亮度）信息与颜色信息分离。在这些颜色模型中，HSI（色相/饱和度/强度）颜色模型非常适合于提供一种更加以人为本的颜色描述方式，因为色相和饱和度成分与人类感知颜色的方式密切相关[ 9]。基于上述对火特征的讨论，可以合理地假设一般火焰的颜色属于红黄色范围。这将映射从0°到60°分布的一般火焰的色相值。如前所述，火的饱和度将随着各种背景照明而变化，即白天获得的饱和度要大于用彩色摄像机捕获可视图像时的夜晚的饱和度。这是因为如果没有其他背景照明，火将成为主要照明。在这种情况下，根据摄像机的操作，火焰的色调将显示为白色。另一方面，当背景照明与火光相当时，图像中的火光颜色在色调上的白色较少。在视频处理中提供足够的亮度。强度应超过某人的阈值。为避免导致火灾，应及早发现燃烧的火焰，并发出火警。尽管有多种颜色的火焰，但初始火焰通常会显示红色到黄色。为了降低计算复杂度，以前的火灾探测算法[7]基于RGB颜色模型，用于从图像中提取火焰区域。相应的RGB值将映射到以下条件：R≥G 和 G > B，即红色到黄色的颜色范围。因此，将要检测的火颜色的条件定义为R≥G>B拍摄的图像中的火灾区域。此外，由于R成为火焰火焰RGB图像中的主要成分，因此在捕获的火焰图像中应该有更强的R。这是因为火也是光源，并且摄像机在夜间需要足够的亮度才能捕获有用的图像序列。因此，R分量的值应超过阈值，[RŤ但是，背景照明可能会影响火焰的饱和度或产生类似火灾的别名，然后导致错误的火灾探测。为了避免受到背景照明的影响，提取的火焰的饱和度值必须超过某个阈值，才能排除其他类似火灾的别名。这将得出三个决策规则[10]，用于从图像中提取火焰像素，如下所述：
To simulate the color sensing properties of the human visual system, RGB color information is usually transformed into a mathematical space that decouples the brightness (or luminance) information from the color information. Among these color models, HSI (hue/saturation/intensity) color model is very suitable for providing a more people-oriented way of describing the colors, because the hue and saturation components are intimately related to the way in which human beings perceive color [9]. Based on the discussion of fire features as described in the above, it is reasonable to assume that the color of general flames belongs to the red-yellow range. This will map the value of hue of general flames to be distributed from 0° to 60°. As mentioned previously, the saturation of a fire will change with various background illuminations that the saturation obtained during the day is larger that that of during the night when the visual image is captured with a color video camera. This is because that the fire will become the major and only illumination if there is no other background illumination. In this situation, the fire-flame will display as more white in the hue according to the operation of cameras. On the other hand, the fire color in an image has less white in the hue when the background illumination is comparable with the fire-light. For providing sufficient brightness in video processing. the intensity should be captured over someone threshold. To avoid leading to a fire disaster, the fire-alarm should be given as soon as detecting a burning fire early, In spite of various colors of fire flames, the initial flame frequently display red-to-yellow color. To reduce computational complexity, the previous fire-detection algorithm [7] is based on RGB color model for extracting the fire region from an image. The corresponding RGB value will be mapped to the conditions: R≥G and G>B, i.e., the color range of red to yellow. Thus, the condition of fire’s colors to be detected is defined as R≥G>B for the fire region in the captured image. Furthermore, there should be a stronger R in the captured fire image due to the fact that R becomes the major component in an RGB image of fire flames. This is because that fire is also a light source and the video camera needs sufficient brightness during the night to capture the useful image sequences. Hence, the value of R component should be over a threshold, RT However, the background illumination may affect the saturation of fire flames or generate a fire-similar alias, and then result in a false fire-detection. To avoid being affected by the background illumination, the saturation value of fire-flame extracted needs to be over someone threshold in order to exclude other fire-similar aliases, This will deduce three decision rules [10] for extracting fire pixels from an image, as described in the following:
在规则 3 ， 表示当R分量的值是 对于相同的像素。根据基本概念，饱和度将随着R分量的增加而降低，因此， 图1示出了当R分量朝着最大值255增加并且然后饱和度将向下减小到零时。对于提取的火像素，R分量和饱和度之间的关系可以绘制在图2中。在决策规则中， 和 根据各种实验结果定义，典型值范围为55至65和115至135 和 ， 分别。
In rule 3,ST denotes the value of saturation when the value of R component is RT for the same pixel. Based on the basic concept, the saturation will degrade with the increasing R component, and thus the term of ((255-R)∗ST/RT) illustrates when R component increases toward the upmost value 255 and then saturation will decrease downward to zero. The relation between R component and saturation for the extracted fire pixels can be plotted in the Figure 2. In the decision rules, both values of RT and ST are defined according to various experimental results, and typical values range from 55 to 65 and 115 to 135 for ST and RT, respectively.
2.2火焰动力学分析
2.2 Dynamics Analysis of Flames
不幸的是，图像中某些类似火的区域可能具有与火相同的颜色，并且通常从图像中提取这些类似火的区域作为真实火。这些火灾别名是由以下两种情况生成的：与火灾具有相同颜色的非火灾对象以及具有类似火灾的光源的照明的背景。在第一种情况下，带红色的物体可能会导致错误地引出火焰。错误提取火灾的第二个原因是，燃烧的火，太阳反射和人造光的照明背景对提取有重要影响，从而使过程复杂且不可靠。
Unfortunately, some fire-like regions in an image may have the same colors as fire, and these fire-similar areas are usually extracted as the real fire from an image. These fire aliases are generated by two cases: non-fire objects with the same colors as fire and background with illumination of fire-like light sources. In the first case, the object with reddish colors may cause a false extraction of fire-flames. The second reason of wrong fire-extraction is that the background with illumination of burning fires, solar reflections, and artificial lights has an important influence on extraction, making the process complex and unreliable.
为了验证真实的燃烧火焰，除了使用色度之外，通常还采用动态特征来区分其他火灾别名。这些火灾动态包括火焰的突然运动，形状变化，增长率和红外响应中的振荡（或振动）。在[1]中。具有增长率的火箱用于检查真实的着火，然后释放指定的动作。另一种方法[5]将失火的程度定义为颜色遮罩后两个连续的最终轮廓图像之间的差异。
To validate a real burning fire, in addition to using chromatics, dynamic features are usually adopted to distinguish other fire aliases. These fire dynamics include sudden movements of flames, changeable shapes, growing rate, and oscillation (or vibrations) in the infrared response. In [1]. fire boxes with the growth rate are used to check a real burning fire and then release a specified action. Another approach [5] defines the degree of fire disorder as the difference between two consecutive final contour images after color masking.
为了提高检测的可靠性，我们利用火焰的无序特性和火象素的增长来检查它是否是真正的火。由于火焰的形状会由于空气的流动而随时变化，因此图像中火区的大小无法保持恒定。并且，就火灾事故而言，火焰总是具有增长的特征。可以用两个连续图像之间的火焰差异的像素量来测量失调。关于失调测量的决策规则推导为：
For improving the reliability of detection, we utilize both the disorder characteristic of flames and the growth of fire pixels to check if it is a real fire. Since the shape of flames is changeable anytime owing to air flowing, the size of a fire’s area in an image can’t maintain to be constant. And, in point of the fire accident, the flame always has a growth feature. The disorder of fires can be measured with the pixel quantity of flame difference between two consecutive images. The decision rule on disorder measurement is deduced as:
哪里 ， 和 分别表示当前和先前的火焰图像。 和 分别表示当前火焰图像和下一个火焰图像的无序值。FTD表示无序阈值，它将与其他类似火的物体区分开。如果满足上述条件（4），则表示火焰可能是真实的火灾，而非火灾。为了提高可靠性，应该执行d次无序检查过程。应该注意的是，FTD和d的两个参数都取决于实验的统计数据。
where , and denotes the current and previous flame image, respectively. and denote the disorder values of current flame image and next flame image, respectively. FTD means a disorder threshold, which will distinguish from other fire-like objects. If the above condition (4) is satisfied, it implies that the flame may be likely a real fire, not fire-alias. For increasing the reliability, the disorder checking process should be performed for d times. It should be noted that both parameters of FTD and d are dependent on the statistical data of experiments.
通常，熊熊燃烧的烈火将主要由气流和燃料类型决定。由于空气流动，火焰大小随时都可以改变，但是它总是朝着越来越近的方向发展，特别是对于最初燃烧的火焰。为了确定火灾的增长特征，我们在每个时间间隔计算一个图像帧的火灾像素数量，然后每两个连续量进行比较。让米一世 和 米我+ 1分别表示当前图像帧和下一个图像帧的火象素数量。如果比较结果米我+ 1> 米一世 超过 G 间隔时间 ŤF 在一段时间内 Ť，在哪里 G，ŤF 和 Ť依靠实验的统计数据。这表明存在可能的火灾增长特征，这将增加真实火灾的验证过程。根据上述从火的色度和动力学分析得出的决策规则，图3和图4显示了分别燃烧纸型燃料和汽油型燃料的火像素。
Generally, the growing of a burning fire will be mainly dominated by the air-flow and fuel type. The flame size is changeable anytime due to air flowing, but it always gets toward the increasing approach, especially for initial burning flame. To identify a fire’s growth feature, we calculate the fire-pixel quantity of one image frame at each time interval and compare every two continuous quantities. Let mi and mi+1 denote the fire-pixel quantities of the current image frame and next image frame, respectively. If the comparing result of mi+1>mi is more than g times at intervals of tF during a time period T, where g,tF and T rely on statistical data of experiments. This reveals that there is a likely fire’s growth feature and this will increase the validating process of a real fire. Based on the above decision rules from the chromatic and dynamics analyses of fire, Figure 3 and 4 show the extractions of fire-pixels for burning the paper-type fuels and gasoline-type fuels, respectively,
