Abstract: Fire detection and early warning technology is an important technical guarantee to effectively reduce fire loss, assist fire fighting and protect people's life and property safety, and smoke detection is one of the more effective means. At present, most smoke detection and alarm devices are mainly set in indoor space, and only have detection and alarm functions. At the same time, the false alarm rate is relatively high, and real-time video information cannot be transmitted synchronously, and detection cannot be carried out for outdoor space areas. Based on the above situation, this paper mainly studies the real-time detection of smoke based on video monitoring system. In the experiment, the network (CNN) architecture was improved, and the res-EfficientNet residual module was added in the EfficientNet, which was used for more accurate detection and recognition of smoke. The spatiotemporal regularized correlation filter (STRCF) was used to achieve the accurate location of smoke. In order to improve the detection accuracy, the polarization transmission characteristics of smoke, such as smoke diffusion and translucency, are also considered. In order to better detect the smoke in the video, the energy of spatial frequency is taken as the one-dimensional constraint term of the filter, and the experiment is carried out on the benchmark data set. The experimental results show that the accuracy is improved by 3%.

Key words: fire smoke, smoke detection, CNN network