室内场景火苗检测算法研究与视觉预警系统设计

发布时间：2018-07-15 12:50

【摘要】：火灾一直威胁着人类生命和财产的安全,提前检测出火灾并发出预警信息一直是一个重要的研究课题。近二十多年来,随着科技的快速发展,基于视觉的火灾检测技术成为预防火灾的一个主要研究方向。基于视觉的火灾检测运用监控视频可以得到更丰富的信息,抗干扰能力更强,适用的区域也更广泛。然而目前基于视觉的火灾检测的研究工作主要是对于已经燃烧起来的火焰进行分析,还没有学者针对初期的火苗进行研究。在公交车、加油站、易燃仓库等场景中,一旦出现火苗,就可能会造成不可估量的损失。根据这些特殊场景的需求,本文针对室内场景的火苗视频,研究了火苗的静态特性与动态特性,提出了一套完整的火苗检测方法。通过不同场景的火苗视频进行测试,本文提出的方法有较低的虚警率和较高的准确率。本文的主要研究内容有:(1)研究了静态火苗的颜色特性与轮廓形状特性。分析了火苗在不同颜色模型下的统计特征,利用火苗的颜色特征分割火苗图像,对比分析火苗图像在不同颜色模型下的分割结果,得到最佳的表示火苗的颜色模型。应用链码得到火苗边界信息,然后用傅里叶描述子、Hu不变矩特征表述火苗的边界形状,对比分析了傅里叶描述子与Hu不变矩特征表述火苗边界的可靠性。(2)研究了火苗视频中的运动目标检测。火苗有其独特的运动特性,火苗出现时可以看作是运动目标,火苗出现后,又是静止不动的。针对火苗运动的特性,提出改进型的帧间差分法检测运动目标,去除大量与火苗颜色相近的背景信息。(3)研究了火苗的帧间动态特征。火苗是一个具有稳定的形态的目标,帧间的火苗具有连续性,火苗在帧间检测出的坐标位置具有相关性。通过对比帧间火苗位置的信息,可以验证火苗判别的正确性。(4)应用支持向量机的分类方式判别火苗是否存在。通过在不同场景拍摄的火苗视频获取大量火苗样本图像与负样本图像,运用支持向量机对正负样本进行训练,然后用测试样本验证火苗判别的准确性,测试的结果表明有较高的识别率。
[Abstract]:Fire has been threatening the safety of human life and property. It is an important research topic to detect fire in advance and issue early warning information. In recent twenty years, with the rapid development of science and technology, visual based fire detection technology has become a major research direction of fire prevention. Visual based fire detection and monitoring video can get more information, stronger anti-jamming ability and more suitable area. However, the research work of fire detection based on vision is mainly focused on the analysis of the flame that has already been burned, and there is no research on the flame in the early stage. In bus, gas station, flammable warehouse and other scenes, fire may cause incalculable damage. According to the requirements of these special scenes, this paper studies the static and dynamic characteristics of the fire, and puts forward a complete fire detection method. The method proposed in this paper has lower false alarm rate and higher accuracy rate. The main contents of this paper are as follows: (1) the color and contour characteristics of static flame are studied. The statistical features of flame under different color models are analyzed, and the flame image is segmented by using the color feature of fire. The best color model is obtained by comparing and analyzing the segmentation results of flame image under different color models. Chain code is used to obtain the flame boundary information, and then the Hu invariant moment feature is used to describe the shape of the flame boundary. The reliability of Fourier descriptor and Hu invariant moment feature in describing fire boundary is analyzed. (2) moving target detection in fire video is studied. Flame has its unique characteristics of motion. When it appears, it can be regarded as a moving target. After it appears, it is still. According to the characteristics of flame motion, an improved inter-frame differential method is proposed to detect the moving target and remove a large amount of background information similar to the flame color. (3) the dynamic characteristics of the flame between frames are studied. The flame is a target with stable shape, the flame between frames has continuity, and the coordinate position of flame detected between frames is correlated. The correctness of flame discrimination can be verified by comparing the information of flame position between frames. (4) support vector machine (SVM) is applied to distinguish the existence of fire. A large number of images of fire samples and negative samples were obtained from different scene videos. Support vector machine (SVM) was used to train positive and negative samples, and then test samples were used to verify the accuracy of fire discrimination. The test results show that there is a high recognition rate.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X932;TP391.41

【参考文献】