Abstract: This paper presents a smoke detection method, which uses the colour, motion orientation and textural features to detect video smoke, and searches the location of fire source at the same time. Firstly, in view of smoke’s colour information, dark channel prior is adopted to extract smoke candidate regions. Secondly, non-smoke regions are excluded by estimating the motion orientation of smoke, as smoke’s motion orientation is not downward. And then, support vector machine is used to classify smoke from smoke candidate regions by extracting a set of smoke features. Finally, the location of fire source is estimated depending on the smoke’s location and the smoke’s frequency of occurrence. Experiments show that the proposed method is effective for smoke detection.

Key words: smoke detection, dark channel prior, motion orientation, fire source

摘要： 提出了一种视频烟雾检测方法，利用烟雾颜色、运动方向以及纹理等特征区分烟雾，并在检测烟雾的同时找到火源位置。引入暗原色先验方法提取出与烟雾颜色相似的区域作为烟雾候选区;通过分析图像局部纹理特征估计图像块的运动方向，排除运动方向向下的非烟雾区域，从而缩小烟雾候选区;将烟雾候选区的一系列特征作为支持向量机的输入，分类为烟雾和非烟雾;根据被检测出的烟雾在视频帧中的具体位置以及对应位置出现烟雾的频数估计火源在视频帧中的位置。与相关算法的实验结果进行了比较，证明了该算法的有效性。

关键词: 烟雾检测, 暗原色先验, 运动方向, 火源