基于视频的火焰检测算法研究

发布时间：2018-05-13 04:22

本文选题：火焰检测 + 特征提取　；参考：《沈阳工业大学》2017年硕士论文

【摘要】：火灾作为自然灾害之一,每天都影响着世界各地的日常生活。传统的火灾探测器有感温、感烟、感光、气敏以及复合式火灾探测器,这些火灾探测器只能对某一特征信息进行采集,受环境影响。随着信号处理、人工智能等技术的发展,产生了基于视频的火灾检测系统。利用摄像设备获取监控场景的影像信息,自动检测火灾并预警,有着成本低、检测率高、响应迅速等优点,相关研究日益受到关注和重视,成为学术研究热点。本文首先总结了国内外火焰检测算法的发展现状及理论研究成果,之后介绍了火焰检测的主流技术。研究并设计火焰视频前景检测方法,分析提取多维度特征有效表征火焰,并实现了基于支持向量机的火焰检测。研究深度学习的方法,并将卷积神经网络用于火焰检测中。在一般的卷积神经网络的基础上,增加了网络深度,实现了对火焰的识别并分析了网络性能。本文主要开展工作如下:(1)提出了一套复杂环境下火焰检测方法。针对目前火焰颜色模型一般为某个颜色区间,无法精确表征火焰颜色特征的问题,在不同空间对火焰颜色进行分析,寻找火焰像素值在空间的分布规律,据此建立更精确的颜色概率分布模型。针对复杂场景下火焰区域提取困难的问题,提出一种基于搜索框的前景提取方法,充分利用火焰的颜色和运动特性,并综合颜色及运动的帧间相关性,准确提取火焰区域。对基于图像块的火焰特征进行研究,融合了火焰颜色显著性特征、空间梯度特征、帧间梯度特征、闪烁特征和火焰质心运动特征等多维度特征,用于火焰的表征和识别。根据融合的特征采用支持向量机进行分类检测,分析了各个特征对分类准确率的影响。(2)考虑到近年来深度学习的快速发展,以及卷积神经网络在图像分类上的卓越表现,将其应用于火焰检测,并与传统的模式识别方法进行比较。设计了具有更多隐层的卷积神经网络结构,实现了对火焰的分类检测,并通过实验分析了卷积神经网络的性能。实验结果表明,本文提出的前景提取算法能够在复杂背景下较精确地提取出目标区块,提升了识别效率。采用设计的多维度火焰特征可以通过支持向量机进行火焰的有效分类识别,效果良好。卷积神经网络能够应用于火焰识别,但需考虑通过提供更多更具有针对性的训练数据集提升泛化能力。
[Abstract]:As one of the natural disasters, fire affects daily life all over the world every day. Traditional fire detectors have temperature, smoke, light, gas and composite fire detectors. These fire detectors can only collect a certain characteristic information and be affected by environment. With the development of signal processing, artificial intelligence and other technologies, a fire detection system based on video has emerged. With the advantages of low cost, high detection rate, rapid response and so on, the use of camera equipment to obtain image information of monitoring scene, automatic fire detection and early warning has become a hot academic research. This paper first summarizes the development and theoretical research results of flame detection algorithms at home and abroad, and then introduces the mainstream flame detection technology. Research and design flame video foreground detection method, analyze and extract multi-dimension feature to represent flame effectively, and realize flame detection based on support vector machine. The method of deep learning is studied, and convolutional neural network is applied to flame detection. Based on the general convolution neural network, the depth of the network is increased, the flame recognition is realized and the network performance is analyzed. The main work of this paper is as follows: 1) A set of flame detection method in complex environment is proposed. In order to solve the problem that the flame color model is usually a certain color interval and can not accurately represent the color characteristics of the flame, the flame color is analyzed in different space to find the distribution of the flame pixel value in the space. Based on this, a more accurate model of color probability distribution is established. Aiming at the difficulty of flame region extraction in complex scene, a method of foreground extraction based on search box is proposed, which makes full use of the color and motion characteristics of flame, synthesizes the correlation between color and motion frame, and extracts the flame region accurately. The flame features based on image blocks are studied. The features of flame color salience, spatial gradient, inter-frame gradient, flicker and flame centroid motion are combined to represent and identify the flame. According to the feature of fusion, support vector machine is used for classification detection. The influence of each feature on classification accuracy is analyzed. The rapid development of deep learning in recent years and the excellent performance of convolution neural network in image classification are considered. It is applied to flame detection and compared with the traditional pattern recognition method. The convolution neural network structure with more hidden layers is designed, and the classification and detection of flame are realized. The performance of convolution neural network is analyzed by experiments. The experimental results show that the proposed foreground extraction algorithm can extract the target blocks accurately in complex background and improve the recognition efficiency. The designed multi-dimensional flame features can be effectively classified and recognized by support vector machine (SVM), and the effect is good. Convolutional neural networks can be applied to flame recognition, but it is necessary to improve generalization ability by providing more and more targeted training data sets.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】