An Unsafe Behavior Detection Method Based on Improved YOLO Framework

In industrial production, accidents caused by the unsafe behavior of operators often bring serious economic losses. Therefore, how to use artificial intelligence technology monitor in a production area real time has become research topic great concern. Based on YOLOv5 framework, this paper proposes an improved YOLO network detect behaviors such as not wearing safety helmets and smoking places. First, proposed uses novel adaptive self-attention embedding (ASAE) model improve backbone reduce loss context information high-level feature map reducing number channels. Second, new weighted pyramid (WFPN) module is used replace original enhanced feature-extraction PANet alleviate too many layers. Finally, experimental results self-constructed dataset show that framework higher detection accuracy than traditional methods. The average increased 3.3%, helmet 3.1%.