| 杨波,张洋,蔡富杰,刘长发,万鹏伟,裴冬.基于时空特征的视频火灾火焰检测技术[J].火灾科学,2025,34(4):298-304. |
| 基于时空特征的视频火灾火焰检测技术 |
| Video fire flame detection technology based on spatiotemporal features |
| |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI：10.3969/j.issn.1004-5309.2025.04.05 |
| 基金项目:中国华电集团有限公司科技项目(CHDKJ23-02-02-36) |
| 作者 | 单位 | | 杨波** | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 | | 张洋 | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 | | 蔡富杰 | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 | | 刘长发 | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 | | 万鹏伟 | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 | | 裴冬 | 四川阿坝金川华电新能源有限公司, 阿坝, 624100 |
|
| 中文关键词: 深度学习 视频火灾火焰检测 空间特征 动静态特征 时序分析 |
| 英文关键词:Deep learning Video fire flame detection Spatial features Dynamic and static features Time series analysis |
| 摘要点击次数: 11 |
| 全文下载次数: 14 |
| 中文摘要: |
| 针对现有视频火灾火焰检测方法在时空域动静态特征提取、多尺度信息融合以及模型复杂度等方面存在的局限与不足而导致检测性能较差的问题,将深度学习卷积神经网络、循环神经网络与传统火灾动静态特征相结合,提出了一种基于时空特征的视频火灾火焰检测方法。方法首先针对火灾空间域特征利用高效率的卷积、池化、激活等操作设计深层卷积网络结构,逐步提取火灾视频帧在不同感受野下的空间特征;然后,引入多尺度自适应融合机制将不同维度空间特征进行加权融合后对火灾区域进行检测,并从火灾面积、质心、分散度等角度对火灾区域动静态信息进行提取;其次,利用循环神经网络从时间域角度对视频序列中的火灾区域特征及动静态特征进行时序分析;最后,结合时域和空间域特征分析结果,实现视频火灾火焰的精确检测。通过在FireFlame、FurgFire和MiviaFire公开数据集上的测试结果有效验证了所提方法各个模块的可行性,与其他视频火灾检测方法相比,该方法展现出更高的准确性和鲁棒性,能够更好地应用于视频火灾检测中。 |
| 英文摘要: |
| In order to solve the problem of poor detection performance caused by the limitations of the existing video fire and flame detection methods in the spatiotemporal domain dynamic and static feature extraction, multi-scale information fusion and model complexity, a video fire flame detection method based on spatiotemporal features was proposed by combining deep learning convolutional neural network, recurrent neural network and traditional fire dynamic and static features. Firstly, the deep convolutional network is designed using efficient convolution, pooling, activation, and other operations to extract spatial-domain features of the fire, and the spatial features of the fire image across different sensing fields are gradually extracted. Then, a multi-scale adaptive fusion mechanism was introduced to detect the fire area by weighted fusion of spatial features of different dimensions, and dynamic and static information of the fire area was extracted from perspectives of area, centroid, and dispersion. Secondly, a recurrent neural network was used to analyse the time series of fire area characteristics and dynamic and static features in the video sequence from a time-domain perspective. Finally, combined with the analysis of time-domain and spatial-domain characteristics, accurate flame detection is achieved. The test results on the public datasets FireFlame, FurgFire, and MiviaFire effectively verify the feasibility of each module of the proposed method. Compared with other video fire detection methods, the proposed method shows higher accuracy and robustness and can be better applied to video fire detection. |
| 关闭 |
|
|
|
