Bi-YOLOv8 Based Safety Detection Algorithm in Power Plant Environment

Timely safety detection is essential to prevent the accidents in the power plants. So far, research in the safety helmets detection has mainly focused on the features such as shape or color. However, it could be severely compromised by various factors, such as numerous facilities, complex shadow occlusions and the presence of smoke and dust. In this paper, a practical safety helmet wearing detection system is proposed, which utilizes a unique backbone network structure, called Bi-YOLOv8 with the improved backbone network structure. A pedestrian safety detection simulation dataset (PPW-dataset) is first formed in the power plant environment. This dataset can not only simulate the real scene of the power plant, but also meet the requirements of common safety detection tasks. Further, a feature extraction method is developed to improve the occluded target recognition by assigning shallow position information to deep features while preserving the original features of the image. Finally, a large number of the experimental results on different public datasets as well as PPW-dataset demonstrate the effectiveness of the proposed object detection method.