在当前社会，火灾是一种最频发、最普遍的灾害之一，它直接危及人类的生命和财产安全。实时高效地预警火灾的发生已经是现代社会存在的重大问题。传统的火灾探测器虽然成本低、见效快，但也存在许多不足，特别是在实时性、检测范围以及鲁棒性上存在问题。随着视频监控设备日益增加以及计算机视觉和图像处理技术的发展，基于视觉特征的视频火灾检测技术已经成为人们研究的焦点。 本文论述了火焰的视觉特性，对比说明了国内外视频火焰检测方法的优势及不足。在此基础上，根据对火焰燃烧过程中所表现的显著性研究，引入计算机视觉领域中的视觉注意机制，构建了一种时空视觉显著性的视频火焰检测模型。模型中主要研究了火焰在空间和时间上的特性，通过仿生视觉注意机制最终实现对火焰目标的检测。 火焰的颜色、亮度以及纹理特性是火焰的空间显著特征。本文在定义火焰颜色显著度时引入了与人类视觉相似的Lab颜色空间，在该颜色空间上采用阈值分割和高斯滤波进行处理，通过差分算法获得火焰的颜色显著图；结合中心-周边差思想与纵向灰度积分投影法来获取亮度显著图；在此基础上将火焰的颜色和亮度显著图进行线性融合，通过分散纹理特性进行判断获得火焰的空间显著图。针对常用火焰运动目标检测方法的不足，提出一种改进的运动目标检测算法。该算法是通过分析多个图像序列来检测运动区域，依据火焰运动累积和颜色特性，并结合背景更新法和帧间差分法的优点来构建火焰的时间显著图。本文在融合火焰的空间、时间显著图之后，通过主运动方向进行最终判断和识别。 为了验证本文算法的有效性，对大量不同场景下的视频进行测试，测试结果表明本文的火焰检测算法能够准确、实时、有效的检测视频中的火焰，且具有一定的抗干扰能力。

In modern time, the fire is one of the most frequent and most common disasters, it directly endangers the safety of human life and property. Constant and effective warning of fire has become a major concern of the society. Traditional fire detectors has low cost and high efficiency, but also has many disadvantages, especially in real-time performance, detection range and the robustness. With the development of video monitoring equipment and the computer vision and image processing technology, the fire detection technology based on visual features has become the focus. In this paper, the visual salient features of the flame were studied, and the advantages and disadvantages of the domestic and foreign video fire detection technology were compared and analyzed. According to a study of the performance in the flame combustion process, the idea of visual outstanding in computer vision was introduced, and the temporal-spatial visual model was integrated. In the model, the characteristics of flame in space and time were studied, and the detection of flame was finally realized by using the bionic visual attention mechanism. Color, brightness and texture are all of spatial salient features of flame. The Lab color space which is similar with human visual was introduced to define flame color saliency and preprocessed by threshold segmentation and Gaussian filter. Then the flame color saliency was obtained through difference algorithm. The brightness saliency map was acquired by combining the center-surround difference and vertical gray-level integral projection method. With the linear fusion of the flame color and intensity saliency , flame spatial saliency map was got by dispersion texture feature. In the article, an improved motion detection algorithm was aimed at making up the shortcoming of common flame moving target detection. According to the new algorithm, the moving regions can be tested on multiple image sequence, and the flame time saliency map can be constructed under the characteristics of movement cumulative and color, combining the advantages of background updating method and frame difference algorithm. Integrated with the time saliency map and spatial saliency map, the flame can be judged and identified ultimately through the main migration orientation. In order to verify the effective of the proposed method, a large number of different scenarios videos were tested. The experimental results proved that the proposed method can detect the flame in video accurately, real-time and effectively, and has strong anti-interference ability.