基于视觉感知机理的林火视频识别模型研究

Due to large forestry area, many disadvantageous factors such as narrow vision and subjective carelessness exist in the process of artifical fire monitoring.Therefore, video surveillance system is widely used in the field of forest fire monitoring, and automatic recognition of forest fire is becoming one of the hot topics in the field of fire warning. But the major difficult problem that fire video surveillance system is how to efficiently detect fire video clip from huge sum of fire video of rapid transmission, so the project plans to study a new fire recognition method for forest fire surveillance syetem. The main characteristics are as follows. The first is the extraction of flame feature vector and smoke feature vector based on motion, temporal and spatial features from forest fire video clip for the purpose of avoiding the influence of environment and weather. The second is the establishment of the forest fire video description model based on ontology. The description model can convert the course of human cognition of fire video into the conceptual framework of ontology description. The third is the establishment of a forest fire recognition model on the basis of visual perception mechanism. The model will detect fire timely and accurately for forest fire video surveillance system, and furthermore, it can be used to retrieve forest fire video clip and establish fire video indexing archives. Therefore the researches in the project will solve urgent problems of forest fire video surveillance systems, and has a broad application prospect.

由于林区范围巨大，人为火情存在视野狭窄，主观麻痹大意等诸多不利因素，因此，森林火灾视频监控系统已广泛应用于林火监测中，火情自动识别也成为森林防火领域的研究热点之一。但如何从数据量巨大的快速传输视频中准确高效地识别火情，仍然是视频监控系统面临的重大难题，因此，本课题拟研究面向森林火灾视频监控系统的高效的火情识别新方法。本课题研究特色主要有：一是拟在视频片段大粒度下,基于时空特征和运动特征提取火焰特征向量和烟雾特征向量，避免气候、环境条件的影响。二是建立林火视频的本体描述模型，将人脑对视频的"认知"过程转换为"本体描述"的概念框架。三是尝试建立一个基于视觉感知机理的林火视频识别模型，及时准确地对林火监控视频进行自动火情识别，而且该模型还可以应用于林火视频片段检索和索引档案的建立。因此，本课题的研究内容是林火无线视频监控系统迫切需要解决的关键问题，具有广阔的应用前景。

本课题主要研究了面向森林火灾视频监控系统的高效的火情识别新方法。主要研究成果：（1）从视频片段大粒度的角度出发，定义了森林火焰视频片段的静态和动态特征的协方差描述子；提出了基于旅行者行为图的森林火焰动态纹理特征提取方法；研究了基于蔓延特征和飘动性特征的森林烟雾视频检测方法；本课题提出的森林火焰和烟雾视频识别方法能够提高林火视频识别的准确率且降低误报率。（2）研究了基于视觉感知机理的图像分类、索引、视频对象检测方法，并应用于森林火灾视频的结构化和描述模型的建立；（3）提出了一种基于语义特征的层次的森林火灾视频识别模型，研究了从像素级到片段级的火灾识别模型的建立方法，建立了森林火灾视觉特征与语义特征之间的桥梁；（4）提出了一种基于非下采样小波变换的LBP纹理特征提取新方法，定义了新的 LBP特征相似度公式，并可应用于更准确地检测森林火灾视频运动区域。