Abstract: Aiming at the low accuracy and poor generalization of the target detection algorithm for safety equipment such as safety helmets, insulating gloves, and insulating shoes of traditional substation electric personnel, especially for the difficulty of detecting whether to wear insulating gloves or not, an improved YOLOv8 detection algorithm Wear-YOLO for substation power personnel safety equipment is proposed. In order to better learn the contextual information of complex scenes, the C2f （CSP bottleneck with 2 convolutions） module of the Backbone part of YOLOv8 is replaced with the MobileViTv3 module that integrates the Transformer structure to capture long-distance dependencies and contextual information and combine it with local information. And the feature extraction capability of the model is improved in substation scenarios. At the same time, in order to optimize the small target detection effect, a small target detection layer is introduced to enhance the  extraction of the network in shallow semantic information, thereby improving the  detection accuracy for small targets not wearing insulating gloves. WIoUv3 is used to improve the bounding box regression loss function, and a dynamic non-monotonic focusing mechanism is introduced to make the model focuses more on ordinary quality anchor boxes, thus improving the accuracy of model detection and its adaptability to complex situations. The experimental results show that the average detection accuracy is 92.1%, the detection accuracy of helmets is 96.8%, the detection accuracy of wearing insulating gloves is 92.1%, the detection accuracy of not wearing insulating gloves is 81.6%, and the detection accuracy of insulating shoes is 93.0%. Compared with the classic target detection models Faster R-CNN, SSD, RetinaNet, and YOLOv5, it has better detection accuracy and robustness.

Key words: safety equipment detection, insulating gloves, YOLO, fusion Transformer, loss function

摘要： 针对传统变电站电力人员的安全帽、绝缘手套、绝缘鞋等安全装备的目标检测算法精度低，泛化性差，尤其针对是否佩戴绝缘手套检测难的问题，提出了一种改进YOLOv8的变电站电力人员安全装备检测算法Wear-YOLO。为了更好地学习复杂场景的语境信息，将YOLOv8的Backbone部分的C2f（CSP bottleneck with 2 convolutions）模块替换为融合Transformer结构的MobileViTv3模块，捕获长距离依赖关系和上下文信息，并与局部信息相融合，提升模型在变电站场景中特征提取的能力。同时为优化小目标检测效果，引入小目标检测层以增强网络对浅层语义信息的提取，从而提升模型对于未佩戴绝缘手套小目标的检测精度。采用WIoUv3改进边界框回归损失函数，引入动态非单调聚焦机制使得模型更专注于普通质量的锚框，从而提高模型检测的准确率和对于复杂情况的适应能力。实验结果表明，平均检测精度92.1%，对安全帽的检测精度96.8%，对佩戴绝缘手套的检测精度92.1%，对未佩戴绝缘手套的检测精度81.6%，对绝缘鞋的检测精度93.0%，并且在与经典目标检测模型Faster R-CNN、SSD、RetinaNet、YOLOv5的对比实验中具有更好的检测精度和鲁棒性。

关键词: 安全装备检测, 绝缘手套, YOLO, 融合Transformer, 损失函数