在智慧型安全監控的應用範疇中，火警偵測一直被視為重要的研究領域，亦即如何在火警初發當下即時偵測火焰與煙霧，以降低生命財產的損失。本文提出了一套以影像處理技術為基礎的火焰煙霧偵測方法：對於火焰，首先建立影像像素梯度在時間軸上的機率密度函數模型，接著以熵的數學原理分析火焰的形變程度，搭配色彩法則作進一步過濾。煙霧方面，第一步使用背景偵測演算法分離出前景及背景影像，並根據煙霧會導致遮蔽區域影像彩度降低的特性從前景中分離出候選區域。其次，因煙霧所到之處會造成背景影像細節模糊化，此現象可透過小波轉換過濾出高頻能量衰減區域。最後使用積分影像對候選區域建立運動方向模型以取得煙霧成份。

In the applications of intelligent video surveillance, fire detection has been regarded as an important field of research. Such security solutions are developed to reduce the loss of lives and properties by detecting flames and smoke in real time. This paper proposes an integrated approach to detect flames and smoke based on computer vision techniques. To detect the existence of flames, firstly Sobel filter is adopted to calculate the gradient value of each pixel and to evaluate the probability density function during a period of time. Then the entropy theory is joined to measure the variation of flame shapes. Flame color rules are also used as a simple criterion to filter out most non-flame pixels. In order to improve the computation efficiency, a lookup table is used for entropy computing. For smoke detection, background of the scene is estimated in advance, and the decrease of high frequency energy is under inspection by applying spatial wavelet transform respectively to the background and the current images. Moreover, the diffusion of smoke also causes grayish and semi-transparent phenomenon of the scene, and these lead to the decrease in chrominance components of image pixels. The smoke regions are finally determined via motion orientation model, which is performed over the candidate blocks using integral images.