检测烟雾可以预警火灾,通过视觉监控烟雾比其他方式监控范围更广,反应更灵敏,对环境要求更低。但是目前的烟雾检测算法,无论是利用单一的色彩、纹理、形状、飘动性、闪烁以及频率等特征,或者要求满足多样特征或者采用支持向量机、神经网络、Bayesian等分类器的方法,都无法保证判断的准确性、适应性和快速性。综述各种烟雾检测的方法后,认为要得到更加可靠的检测效果,一方面需要更加本质的烟雾特征,另一方面要对已有的特征进行更深入的实验验证,同时也要有更全面的样本视频数据库和算法评价标准。

Smoke detection with no latency for fire alarming is crucial to minimize fire damages and saving lives. Video as a spatio-temporal sensor covers a larger area than point sensors and it is sensitive to environment changes. Current smoke detection algorithms, however, are still difficult to achieve fast, accurate, and robust judgment on fire, even though they have used chromatic characters, texture, shape, flutter, flicker, spatial and temporal frequencies, as well as composite classifiers such as support vector machine, neural network, etc. This survey reviews the state-of-art of smoke detection methods and proposes three directions to achieve robust data processing. We argue that the success of visual smoke detection should be consolidated by a rigorous understanding of its physical characteristics, creation of a common test database for algorithm validation and comparison, and the establishment of a new criteria for the evaluation of algorithms.