Abstract:

Most existing video smoke detection methods identify candidate smoke image patches by motion detection, and manually extract combined features of smoke, which accuracy is difficult to meet requirements in complex scenes. Aiming at the above problems, a video smoke detection method based on a spatiotemporal dual path 3D residual convolutional network is proposed. The candidate smoke regions are extracted based on the wavelet low-frequency component difference between the mixed Gaussian background model and the original video frame. A spatiotemporal dual path 3D residual convolutional network is constructed, and the attention module is introduced to the weighted fusion of spatial and temporal features of smoke to achieve end-to-end smoke recognition. The experimental results show that this method can get a more complete candidate smoke regions, especially for the smoke which is too thin or thick, and compared with the traditional smoke detection method and 2D smoke detection convolutional network, smoke detection accuracy has been improved.

Key words: 3D residual convolutional network, smoke detection, attention module, deep learning

摘要：

现有的视频烟雾检测方法大多通过运动检测提取疑似烟区，并依据经验手工设计提取烟雾特征，在复杂场景中检测准确率不高。针对以上问题，提出了一种基于时空双路3D残差卷积网络的视频烟雾检测方法，基于混合高斯背景模型与原始视频帧的小波低频分量差进行疑似烟区提取，其次构造时空双路3D残差卷积神经网络，并引入注意力机制加权融合烟雾时空域特征，实现端对端的烟雾识别。实验结果表明，该方法可以得到更为完整的疑似烟区，尤其对于过于稀薄和浓厚的烟雾分割效果较好，且相比于传统的烟雾检测方法和2D的烟雾检测卷积网络，在烟雾检测准确率上得到了提高。

关键词: 3D残差卷积网络, 烟雾检测, 注意力机制, 深度学习