Innovative YOLOv5s-Based Algorithm for Real-Time Tunnel Fire Detection

Aiming at the special characteristics of tunnel spatial environment and the problems of confusing smoke and fire in tunnel fires with high real-time detection requirements, an improved algorithm for tunnel smoke and fire detection based on YOLOv5s is proposed. The Focus module is replaced by a convolutional layer, the number of convolutions in B_CSP is reduced, and the spatial pyramid pooling structure SPP is replaced by SPPF, which reduces the number of parameters in the network and improves the detection efficiency of the model. In order to better optimize the anchoring frame, the dynamic, non-monotonically focused Wise-IOU bounding box loss function is used to replace the GIOU, which speeds up the convergence of the network and improves the accuracy of network detection. The CA attention mechanism is incorporated at the end of the Backbone layer to dynamically adjust the importance between channels, which enables the model to better capture the flame features in the image and improve the model performance. The experiments use 2050 tunnel smoke and fire datasets as training samples, and the experimental results show that the accuracy of the improved model on the dataset is increased by 3.2 percentage points, and the network model can achieve 98.9% precision and 95.1% recall as well as 99.2% average precision by testing the network model on the dataset, and the detection speed is increased to 148 FPS. It can meet the real-time prevention and detection of tunnel fire in daytime, nighttime or poor vision, and has good accuracy and robustness.