基于改进层次聚类和SVM的图像型火焰识别

摘要

为了提高大空间建筑内实时监控的火灾检出率，提出基于改进分层聚类和支持向量机（SVM）的火灾识别算法。首先建立火焰颜色模型，用像素运动累积法获取疑似目标，借助改进层次聚类法对其进行合并，形成少量疑似区域。然后提取疑似区域相邻帧间相关性、面积变化率、质心偏移距离、红绿分量比、平均亮度这五个特征量。最后将特征输入到SVM进行二分类，判断是否有火。实验结果表明该算法提高了聚类算法在实际应用中的效率，克服了已有火灾识别算法过分依赖阈值的局限性，适用于室内大空间基于视频监控的火灾探测。%In order to improve the fire detection rate based on video monitoring in spacious buildings, a fire detecting algorithm based on improved hierarchical cluster and Support Vector Machines(SVM)is proposed. Firstly suspected tar-gets are detected with pixel motion accumulating method after color detection with a proper color model and the targets number is reduced with an improved hierarchical cluster method. Then the features, inter-frame correlation, area rate, cen-troid offset, average brightness, proportion of green and red are extracted. Finally the features are entered into the SVM to make a decision. The experimental results show that the cluster efficiency is improved, the limitation of threshold depen-dence is overcome, and it is suitable for image fire detection in spacious buildings.