基于时空特征的林火视频火焰识别研究

由于森林火灾监控现场是野外广阔的林区,而且树叶的摇摆使得视频图像中的运动对象很多,强烈的阳光、秋季的枯叶和红枫会导致火灾识别的误报等,这些特点都使得现有室内或静止场景火灾视频监控的视频处理技术不再适用。考虑到火灾发生有一个蔓延的过程,是一个包含若干连续视频帧图像的视频片段,笔者首先将疑似火焰视频划分成时空视频块,根据颜色特征和运动特征得到疑似火焰区域,然后在视频片段大粒度下基于空间静态特征(纹理、圆形度特征)和时序动态特征(火焰面积变化、形状相似性、闪烁频率特征)提取火焰特征向量,最后使用基于AdaBoost的算法进行火焰识别,实现森林火灾的实时检测。结果表明,该方法能够准确有效地进行林火视频火焰识别。
The forest surveillance environment is a large wilderness space, where the swing of leaves causes a lot of moving objects in video images, and strong sunlight, yellow leaves in autumn and red maple leaves cause false fire alert. These features make current fire video detection methods originally used for indoor or static condition unsuitable for forest fire detection. In this paper, flame feature vectors are extracted based on static features and dynamic features of video clips with large granularity, and recognition of forest fire based on AdaBoost algorithm is proposed. First of all, the objects in forest scene are mostly green trees or grass, and the flame color at the early stage is red to yellow. Therefore, color features can be used to exclude a large number of non-fire videos. However, some flame-color objects such as yellow leaves in autumn, red maple leaves and lamplight in woodlands can't be filtered only by color features. Since the shape of flame changes constantly, dynamic features are used to further remove some relatively static flame-alike objects. Secondly, there is a spreading process after fire breaks out, and fire video clips contain a number of consecutive video frames. So forest fire surveillance videos are divided into spatio-temporal video blocks according to sliding time window. Static, dynamic and spatio-temporal features are analyzed based on video clips with large granularity. This paper mainly analyzes two static features(texture features, circularity)and three dynamic features(flame area variation, shape similarity, flicker frequency)to extract flame feature vectors. Finally, considering the shortcomings of conventional flame classification algorithm, this paper presents a flame video recognition model based on AdaBoost algorithm, which can produce a strong classifier by combining a set of weak classifiers―BP neural network. The experimental results showed that the true recognition rate could be as high as 100%, and that the proposed approach could accurately and effectively recognize forest fire flame