Abstract:

To improve the detection rate and accuracy of initial fire in highway tunnel, considering the characteristics of the small amount of initial fireworks and difficult detection, an improved fire detection method based on YOLOv5s was proposed. The "transformer prediction heads" was incorporated into the YOLOv5s feature detection layer, and a fourth 160×160 scale feature detection head was added on the basis of the original three feature detection heads to enhance the multi-scale recognition capability. Meanwhile, the bidirectional feature pyramid bidirectional feature pyramid network(BiFPN)structure was introduced to blend the semantic information of high and low layers of flame and smoke. Complete intersection over union(CIoU) was used to replace distance intersection over union (DIoU), and focal loss was used in confidence loss to improve YOLOv5s loss function, so as to improve the overall training effect and detection accuracy of the new model. The experiment was carried out to simulate the initial fire in a real tunnel, and 50 000 training set samples were obtained. Combined with the video data of the real fire in Guyinshan tunnel in Zhenjiang, Jiangsu province on March 1, 2022, the YOLOv5s-Opt and YOLOv5s algorithm models were compared and analyzed. The results show that the average detection accuracy of YOLOv5s-Opt is 90.38%, which is 2.06% higher than that of YOLOv5s. For the same fire measurement video, the detection rate of YOLOv5s-Opt is 3.63% higher than that of YOLOv5s. The YOLOv5s-Opt algorithm model performs better at the detection and identification of small fire targets in the initial stage, and has advantages in detection accuracy and detection rate, and the detection effect is obvious, which can meet the needs of actual fire detection.

Key words: YOLOv5s algorithm, highway tunnel, initial fire, object detection, deep learning