Xing Jian, Zhan Chenglong, Ma Jiaqiang, Chao Zibo, Liu Ying
School of Electronic Information Technology, Northeast Forestry University, Harbin, 150040, People's Republic of China.
Sci Rep. 2025 Jan 7;15(1):1227. doi: 10.1038/s41598-024-83391-7.
To address the significantly elevated safety risks associated with construction workers' improper use of helmets and reflective clothing, we propose an enhanced YOLOv8 model tailored for safety wear detection. Firstly, this study introduces the P2 detection layer within the YOLOv8 architecture, which substantially enriches semantic feature representation. Additionally, a lightweight Ghost module is integrated to replace the original backbone of YOLOv8, thereby reducing the parameter count and computational burden. Moreover, we incorporate a Dynamic Head (Dyhead) that employs an attention mechanism to effectively extract features and spatial location information critical for site safety wear detection. This adaptation significantly enhances the model's representational power without adding computational overhead. Furthermore, we adopt an Exponential Moving Average (EMA) SlideLoss function, which not only boosts accuracy but also ensures the stability of our safety wear detection model's performance. Comparative evaluation of the experimental results indicates that our proposed model achieves a 6.2% improvement in mean Average Precision (mAP) compared to the baseline YOLOv8 model, while also increasing the detection speed by 55.88% in terms of frames per second (FPS).
为解决建筑工人不当使用头盔和反光服所带来的显著增加的安全风险，我们提出了一种专为安全着装检测量身定制的增强型YOLOv8模型。首先，本研究在YOLOv8架构中引入了P2检测层，这极大地丰富了语义特征表示。此外，集成了一个轻量级Ghost模块来取代YOLOv8的原始主干，从而减少参数数量和计算负担。而且，我们加入了一个动态头部（Dyhead），它采用注意力机制来有效提取对现场安全着装检测至关重要的特征和空间位置信息。这种调整在不增加计算开销的情况下显著增强了模型的表征能力。此外，我们采用了指数移动平均（EMA）滑动损失函数，这不仅提高了准确率，还确保了我们的安全着装检测模型性能的稳定性。实验结果的对比评估表明，与基线YOLOv8模型相比，我们提出的模型在平均精度均值（mAP）上提高了6.2%，同时在每秒帧数（FPS）方面检测速度提高了55.88%。
