基于随机森林特征选择的视频烟雾检测

Features selection for video smoke detection using random forest

利用随机森林算法,提出了一种基于随机森林特征选择的视频烟雾检测方法.首先,提取四种表征烟雾的特征:RGB颜色特征,小波变换高频子图,多尺度局部最大饱和度,多尺度暗通道;其次,根据烟雾图像信息模型利用无烟图片合成烟雾图片并分块得到随机森林训练样本;第三,训练随机森林进行特征选择并通过训练支持向量机得到识别烟雾块和非烟雾块的分类器,并由此得到视频图像帧的疑似烟雾区域;最后通过视频烟雾区域的凸形度和增长率分析,得到烟雾检测的结果。实验结果表明,该方法能够及时的预警烟雾同时降低火灾预警的误报率.

Using the random forest algorithm, a video smoke detection method with features selection was proposed. The method first extracted four original smoke image features including color features in RGB space, wavelet high frequency sub-images, multi-scale local max saturation, and multi-scale dark channel to input the random forest（RF）. Then it utilized haze image formation model to make the synthetic smoke images from non-smoke images and partitions these images into blocks as the samples for RF. Thirdly, it trained RF to get the selected features from the original features and used support vector machine （SVM） to get a classifier which recognizes the smoke blocks and the non-smoke blocks. And then the smoke region candidate can be extracted from video images by the classifier. Finally, the method analyzed the detected smoke region with the features of the growth rate and the perimeter to area ratio to make the final decision on video smoke detection. The experimental results show that the proposed method can detect the smoke timely and give a fire alarm with a lower false-alarm rate.