Tunnel Fire Detection based on Improved YOLOv5

Aiming at solving the problem of detecting tunnel fire precisely and timely, a tunnel fire detection approach based on improved YOLOv5 is proposed in this paper. At first, a multi YOLOv5 models is designed, in which each sub-model is trained by a different subset of data, and makes the multi models have good performance over the certain given environment. Furthermore, CBAM attention mechanism module and RFB module are introduced to improve the performance of YOLOv5 model through special feature extraction networks, which can extract extra abstract feature from feature maps. Compared with the original YOLOv5, the average accuracy of flame detection over 2303 flame images is increased from 89.4% to 92.0%, the recall rate increased from 85.8% to 88.4% and the average speed of detection reached 55.9FPS. The result shows that the optimized tunnel flame detection algorithm has high accuracy with real-time performance, makes it more suitable for tunnel fire detection.