基于视频的火灾烟雾检测算法的研究

发布时间：2018-04-26 11:22

  本文选题：背景动态更新 + 暗通道先验　； 参考：《华侨大学》2017年硕士论文

【摘要】：火灾是严重危害人类生命财产安全和自然生态环境的重大灾害之一。火灾的及时预警对于减少各项损失意义重大。一般火灾发生初期火焰较小,但是烟雾却很明显,因此对火灾烟雾的检测是及时判断火灾是否发生的重要依据。传统的火灾检测技术依赖传感器工作,在开放空间中的使用受到限制。随着智能监控设备的普及,基于视频图像的火灾烟雾检测技术受到广泛关注,它可以有效避免部分环境因素产生的影响,且在大尺度空间监测上具有明显的优势。本文提出了一种动态检测和静态分类相结合的基于视频图像的火灾烟雾检测方法,分别从候选烟雾区域提取和图像特征提取与分类等方面重点研究了火灾烟雾检测算法,主要的工作包括:(1)提出一种基于背景动态更新与暗通道先验的火灾烟雾检测算法。首先通过改进的背景动态更新算法提取运动前景,解决了传统运动目标检测算法针对扩散缓慢的烟雾做前景检测时,容易出现的空洞现象;然后针对目前算法在复杂环境下适应性不强的问题,例如自然场景中存在着诸如树枝晃动、行人和其他运动物体的干扰,很容易产生误检,提出一种基于暗通道先验知识的干扰物体过滤方法,该方法结合运动目标检测算法,使其在候选烟雾区域提取阶段就可以消除多数干扰物体;最后通过多特征融合的方式实现分类识别。实验结果表明算法可以有效减少误检,提升检测性能。(2)提出一种基于卷积神经网络的火灾烟雾检测算法。由于烟雾没有固定的颜色和轮廓,传统基于手工设计特征的烟雾检测算法难以描述烟雾的本质属性,进而影响检测的准确性;同时手工设计和处理特征需要一定的专业知识和经验,这些因素给火灾烟雾检测研究带来难度。因此,本文在前面研究的基础上提出一种基于卷积神经网络的火灾烟雾检测算法,算法通过多层的网络结构能够自动地学习更具判别性的高层特征。高层特征使得算法对于目标的表观变化具有一定的鲁棒性,适合烟雾这类变化物体的特征提取。实验表明通过采取的隐含扩大候选区域策略结合深度卷积神经网络强大的特征抽取能力,大大提高了视频烟雾检测的准确性和及时性。
[Abstract]:Fire is one of the major hazards that seriously endangers the safety of human life and property and natural ecological environment. The timely warning of fire is of great significance to reduce the loss of all kinds of losses. The fire is very small at the beginning of the fire, but the smoke is very obvious. So the detection of fire smoke is an important basis for timely judgment of the occurrence of fire. Disaster detection technology relies on sensor work and is limited in the use of open space. With the popularization of intelligent monitoring equipment, fire smoke detection technology based on video image has been widely concerned. It can effectively avoid the impact of some environmental factors and has obvious advantages on large scale space monitoring. A fire smoke detection method based on video image based on dynamic detection and static classification. The fire smoke detection algorithms are studied from the extraction of candidate smoke region and image feature extraction and classification. The main work includes: (1) a fire smoke based on the background dynamic update and the dark channel prior is proposed. Detection algorithm. Firstly, the motion foreground is extracted by the improved background dynamic updating algorithm, which solves the cavitation phenomenon which is easy to appear when the traditional moving target detection algorithm is easy to detect in the foreground detection of the slow diffused smoke, and then to the problem of poor adaptability in the complex environment, such as the natural scene, such as the branch. The interference of sloshing, pedestrians and other moving objects is easy to be misdiagnosed. A filtering method based on the prior knowledge of dark channels is proposed. This method combines the moving target detection algorithm to eliminate most of the interference objects in the extraction phase of the candidate smoke region; finally, the classification recognition is realized by multi feature fusion. The experimental results show that the algorithm can effectively reduce the error detection and improve the detection performance. (2) a fire smoke detection algorithm based on convolution neural network is proposed. Because the smoke does not have fixed color and contour, the traditional smoke detection algorithm based on manual design features is difficult to describe the essential properties of the smoke and then affects the accuracy of the detection. At the same time, the manual design and processing features require certain professional knowledge and experience. These factors bring difficulty to the fire smoke detection and research. Therefore, a fire smoke detection algorithm based on the convolution neural network is proposed on the basis of the previous research. The algorithm can learn more discriminant by the multi-layer network structure. The high level feature makes the algorithm have a certain robustness to the apparent change of the target, and is suitable for the feature extraction of the change objects such as smoke. The experiment shows that the accuracy of the video smoke detection is greatly improved by combining the implicit extended candidate region strategy with the powerful feature extraction ability of the deep convolution neural network. Timeliness.

【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X932;TP391.41
，

本文编号：1805812