Personnel Safety Equipment Wearing Detection Technology Based on Improved YOLOv8 for Energy Storage Station

Personnel Safety Equipment Wearing Detection Technology Based on Improved YOLOv8 for Energy Storage Station

연구 분야: Infrastructure

학회: CAIBDA '24: Proceedings of the 2024 4th International Conference on Artificial Intelligence, Big Data and Algorithms

The power industry is developing rapidly, the importance of new energy storage technology has become increasingly prominent. Among them, personnel safety is the cornerstone of keeping the power station running steadily, and the staff must wear safety equipment correctly to ensure their own safety and the stable operation of the equipment. This thesis proposes an improved YOLOv8 algorithm for the detection of personnel safety equipment in energy storage power stations, such as helmets, safety belts, insulated gloves, etc., in view of the low detection accuracy and the difficulty in detecting small targets of the detection algorithms for personnel safety equipment wearing in energy storage power stations. First of all, the feature fusion part of YOLOv8 is replaced by Bifpn module, by introducing top-down and bottom-up bidirectional connections, feature information at different scales is fully exploited to improve target detection accuracy. Secondly, the RFB-s module is positioned in front of the detection head to use convolution kernels with varying expansion rates. This enhances the ability to extract finer and richer feature information and increases the accuracy of detecting small targets. Finally, in order to describe the relationship between boundary boxes more accurately, the WIoU loss function is introduced. In the calculation of boundary box loss, not only the intersection ratio is considered, but also the weight of geometric information such as the aspect ratio is added, so as to describe the relationship between boundary boxes more accurately, improve the effect of detection. The experimental results show that the average detection accuracy is 86.6%, which is increased by 1.7% compared to the YOLOv8 algorithm. The detection accuracy of the safety helmet is 88.3%, the insulation glove is 89.2%, and the safety belt is 82.3%. Moreover, compared with the classical target detection models YOLOv3, YOLOv5 and YOLOv6, it has better detection accuracy and robustness.