Optimized YOLOv5 model for safety helmet and flame detection system

Abstract As industrial and safety standards improve, target detection technology is crucial for workplace safety and fire monitoring. The aim of this paper is to enhance the YOLOv5 model for efficient safety helmet and flame detection. The current YOLOv5 faces challenges with class imbalance, small target detection, and robustness in complex environments. This study will optimize YOLOv5 by modifying its backbone network using open-source and self-collected datasets, and incorporating bidirectional feature pyramid network (BiFPN) and spatial and channel self-attention mechanisms. These improvements aim to increase the sensitivity and accuracy of small target detection and enhance robustness under varying conditions. The real-time detection system integrates the optimized YOLOv5 for accurate safety helmet and flame detection in video streams, improving detection confidence value and reliability of small object detection. Furthermore, this paper proposes a comparison of different YOLOv5 models and uses a self-integrated dataset to obtain the optimized model for this system.