Real-Time Video Fire Detection via Modified YOLOv5 Network Model

Accidental fire outbreak threatens people's life and property safety, and it is of great significance to study fire detection and alarm early. The detection range of traditional fire detectors is limited, and the conventional detection algorithm has the problems of low precision and long detection time. Aiming at these problems, a video fire detection method based on improved YOLOv5 is proposed in this paper. To improve the ability of feature extraction and small-scale target detection, the dilated convolution module is introduced into the SPP module of YOLOv5, the activation function GELU and the prediction bounding box suppression DIoU-NMS are employed in the structure of the improved YOLOv5. The experimental results show that the algorithm has fast detection speed and high detection accuracy. It can accurately detect not only large-scale flame but also small-scale flame in the early stage of fire. The precision and recall of the improved small YOLOv5 are 0.983 and 0.992, the mAP@.5 is as high as 0.993, and the detection speed reaches 125 FPS. The proposed method can well suppress false detection and missed detection in complex lighting environments and improve the robustness and reliability of fire detection, meet the performance requirements of the video fire detection task.

Automated summary

This paper proposes a video fire detection method based on improved YOLOv5. By introducing the dilated convolution module and improving the structure design, it enhances the ability of feature extraction and small-scale target detection. The experimental results show that it has a fast detection speed and high detection accuracy.