Fire is one of the disasters with the highest probability among natural disasters and social disasters. It poses a serious threat to human life and life safety. In order to reduce fire losses, a reliable fire warning method is particularly important. But due to huge variations of smoke in color, shapes, and texture and complex application environments, the existing methods still do not meet the application requirements well. To solve these problems, in this paper, we propose a two-stage real-time video smoke detection method base on dense optical flow and convolutional neural network. In the first stage, we propose a fast pre-positioning module to obtain suspicious smoke areas through the dynamic characteristics of smoke which can greatly reduce the subsequent computational complexity, and only extract the moving optical flow of suspicious smoke areas as the dynamic features of the smoke which reduce the subsequent processing time cost. Instead of simply using moving optical flow as the dynamic characteristics of smoke, we found that the optical flow of the blue channel (OFBC) can effectively reflect the motion characteristics of smoke, so we combine the OFBC of suspicious smoke areas with its three RGB color channels to form a quaternion matrix for subsequent classification. In the second stage, we choose ResNet as our pre-classifier, and a temporal enhanced adjustment algorithm was proposed as the pre-classified follow-up fine optimization module, which can fully utilize the characteristics of the smoke movement in the video to improve detection rate. The experimental results show that compared with the existing smoke detection methods, our proposed method achieves high detection rate and low false alarm rate.

火灾是自然灾害和社会灾害中可能性最高的灾害之一。它对人类生命和生命安全构成了严重威胁。为了减少火灾损失，可靠的火灾预警方法尤为重要。但是由于烟雾在颜色，形状和质地上的巨大差异以及复杂的应用环境，现有的方法仍然不能很好地满足应用需求。为了解决这些问题，本文提出了一种基于密集光流和卷积神经网络的两阶段实时视频烟雾检测方法。在第一阶段，我们提出了一种快速的预先定位模块，可通过烟雾的动态特性来获取可疑烟雾区域，从而可以大大降低后续的计算复杂度，并且仅提取可疑烟雾区域的移动光流作为烟雾的动态特征，从而减少了后续处理时间。我们发现蓝色通道（OFBC）的光流可以有效地反映烟雾的运动特征，而不是简单地使用移动光流作为烟雾的动态特征，因此我们将可疑烟雾区域的OFBC与它的三种RGB颜色结合在一起通道形成四元数矩阵以进行后续分类。在第二阶段，我们选择ResNet作为预分类器，并提出了时间增强调整算法作为预分类的后续精细优化模块，该模块可以充分利用视频中烟气运动的特征来提高检测效率。率。