基于时间平滑多特征量火灾快速识别算法

Fast Recognition Algorithm Research of Fire Flame Based on Multifeatures Longitude Regression with Temporal Smoothing

为了降低工矿企业下的火灾识别的真负率,在原有多特征量对数回归识别算法的基础上加上了时间平滑算法。根据火焰的色度特征进行图像分割,通过运动目标与参考图像差分运算获取火焰候选区域,然后提取候选区域的面积变化率、圆形度、尖角个数以及质心位移等特征量,建立火焰的快速时间平滑对数回归快速识别模型。通过已有火焰识别数据库、实验室自制火焰视频进行算法的训练和检验。测试结果表明,与其他已有方法相比,该算法的真正率(TPR)达到90%,真负率(TNR)达到97%,效果较好,为煤工矿企业下实时火灾预警提供了理论基础。

A fast flame recognition algorithm based on multi-features longitude regression with temporal smoothing is proposed to reduce the true negative rate in surveillance video of the coal mine.The image frames are segmented according to the chrominance features of fire flame. The flame candidate region(FCR) is acquired by subtraction the moving target image with reference image. So, the multi features of the FCR such as the changed rate of area, circularity, the number of sharp corners,centroid motion, and so on are extracted to conduct a fast longitude regression recognition model with temporal smoothing. The algorithm is trained and validated by using the fire flame recognition databases and our self-made fire flame videos. The experimental results indicate that the true positive rate(TPR)and true negative rate(TNR) are 90% and 97% compared with the other methods, respectively. So the proposed approach has provided some theoretical bases in the next real-time fire flame detection surveillance video in the coal mine.