In order to overcome the disadvantages of traditional fire detection,such as low sensitivity and speed,this paper proposed an image fire detection method based on HOFHOG bag-of-features and RF.It probed characteristic description by HOF and HOG as flame and smog’s spatial-temporal features.It proposed an analysis way of histograms of oriented optical flow(HOF)in different channels.It probed how to describe construct fire region’s HOFHOG visual dictionary by K-means. It presented a way of adopting block’s frame difference statistic attributes as spatial-temporal fusion feature.It probed parame-ter selection and performance analysis for random decision forest classifier training based on feature subsets via relief feature selection.It detected flame and smog region simultaneously and submitted a more logical alarm judgment by decision tree forest voting according to detected region’s spatial-temporal distribution and relations.The experiment shows that the flame and smog detection system based on HOFHOG bag-of-features and random decision forest classifier has stable higher accuracy.%针对传统火灾探测中灵敏度不高、反应慢的问题,提出一种基于HOFHOG特征词袋和RF的火灾区域探测。探索用光流直方图和有向梯度直方图描述火焰和烟雾的时空特征,提出在时空块内对不同通道下的光流直方图的分析,探索火灾区域的梯度方向直方图的静动态特征的描述方法,将 HOFHOG和其他特征通过 K-means方法构成特征词典,并对随机决策森林树训练过程中的参数、性能进行了选择和分析,同时探测了火焰与烟雾区域各特征的空间分布和时序关系,并由决策森林投票给出逻辑更合理的判断。实验证明,基于HOF与HOG等特征词袋和随机决策森林结合的分类方法在火灾探测系统中表现出了稳定的识别精度。
展开▼