A Dual-Branch Network with Motion Foreground Perception for Tunnel Smoke Detection

As is well known, tunnel fires pose a serious threat to the lives of drivers and passengers. Iidentifying smoke, which is a key element in early fires, is crucial for enhancing tunnel fire prevention. Current vision-based smoke detection method generally exhibits problems such as low robustness, missed detections and difficulty in deployment. Additionally, existing research lacks publicly available tunnel smoke datasets for model training. In this paper, a dual-branch network architecture is proposed and a video dataset of tunnel smoke is constructed by simulating smoke in actual tunnels. The proposed network utilizes Adaptive Gaussian Mixture Model (AGMM) to extract motion foreground mask as auxiliary branch input, which helps the network to perceive subtle motion of smoke. A 3D-Activation Gated Fusion Unit (3D-AGFU) is proposed to efficiently fuse spatial features of the moving foreground with the static semantic features of RGB video frame. Experimental results of testing on self-built tunnel fire datasets show that the Precision, Recall and of the proposed method are 96.4%, 93.5% and 95.9%, respectively. The proposed method significantly improves the accuracy and reduces missed detections of small and thin smoke. Moreover, the proposed network maintains a lightweight design, making it more deployment-friendly. The proposed method is applied to the tunnel smoke detection system developed in this paper, which provides real-time warning function for the management.