基于目标跟踪与深度学习的视频火焰识别方法

Video Flame Detection Method Based on Target Tracking and Deep Learning

摘要:

近年来火灾事故频发,对生态环境,社会经济都造成了严重影响,视频监控系统在火灾预防和环境监控中都有非常重要的作用。针对传统的视频火焰检测方法需要手工提取火焰特征且识别率低、误检率高的缺点,提出了一种基于特征检测,多目标跟踪和深度学习的火焰检测算法。通过高斯混合模型运动检测方法对视频中的动态目标进行提取,再经过HSI与RGB结合的颜色模型进行筛选,得到疑似火焰目标,对提取的目标进行多目标跟踪算法跟踪,最终对稳定存在的目标通过CaffeNet模型进行判别,得到火焰判别结果。实验证明,本算法实现了对视频火焰的准确检测,能对火焰进行有效识别,对火焰视频数据集上的平均识别精度达到98.79%,并能适应实时检测火灾的需求。

Abstract:

In recent years, frequent fire accidents have seriously affected the ecological environment and social economy,video surveillance system plays a very important role in fire prevention and environmental monitoring. To overcome the shortcomings of traditional video flame detection methods, such as manual extraction of flame features, low recognition rate and high false detection rate, a flame detection algorithm based on feature detection, multi-target tracking and deep learning is proposed. The algorithm extracts the dynamic object in video by the Gaussian mixture model motion detection method, then filters the suspected flame object through the color model combined with HSI and RGB, and tracks the extracted object with multi-target tracking algorithm. Finally, the stable target is judged by CaffeNet model and the result of flame discrimination is obtained. Experiments show that the algorithm can detect the video flame accurately, identify, and the average recognition accuracy of the flame video data set is 98.79%, which can supply the demand of real-time flame detection.