基于HOFHOG特征词袋和RF的火灾区域探测  

Detecting fire region based on HOFHOG bag-of-features and RF

作　　者：蒋先刚[1] 范自柱[1] 张盼盼[1] Jiang Xiangang;Fan Zizhu;Zhang Panpan(School of Science, East China Jiaotong University, Nanchang 330013 , China)

机构地区：[1]华东交通大学理学院，南昌330013

出　　处：《计算机应用研究》2016年第10期3160-3164,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61262031;61263032)

摘　　要：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Kmeans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。In order to overcome the disadvantages of traditional fire detection , such as low se n sitivity and speed, this paperproposed an image fire detection m ethod based on HOFHOG bag-of-features and RF . I t probed cha racte ristic d e scription by HOF and HOG as flam e and sm og5 s spa tial-tem poral features. I t proposed an analysis way o f histogram s o f oriented op ticalflo w (H O F ) in d iffe re n t channels. I t probed how to describe construct fire region 5 s HOFHOG visual dictionary by K-means.I t presented a way o f adopting b lo c k 5 s fram e difference statistic attribute s as spatial-tem poral fusion feature. I t probed param eterselection and perform ance analysis fo r random decision forest classifie r train in g based on feature subsets via re lie f featureselectio n. I t detected flam e and smog region sim ultaneously and subm itted a more lo g ica l alarm judg m en t by decision tree forestvotin g according to detected region 5 s spatial-tem poral distribution and relations. The experim ent shows that the flame and smogdetection system based on HOFHOG bag-of-features and random decision forest cla ssifie r has stable high er accuracy.

关 键 词：火灾探测 光流直方图 特征词袋 随机森林 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]