计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (8): 320-328.DOI: 10.3778/j.issn.1002-8331.2301-0057

• 工程与应用 • 上一篇    下一篇

自然场景下配电线网施工安全帽佩戴检测算法

许逵,李鑫卓,张历,张俊杰,杨宁   

  1. 1.贵州电网有限责任公司 电力科学研究院,贵阳 550002
    2.武汉光谷信息技术股份有限公司,武汉 510220
  • 出版日期:2024-04-15 发布日期:2024-04-15

Safety Helmet Wearing Detection Algorithm for Distribution Network Construction in Natural Scenarios

XU Kui, LI Xinzhuo, ZHANG Li, ZHANG Junjie, YANG Ning   

  1. 1.Electric Power Research Institute of Guizhou Power Grid Co., Ltd., Guiyang 550002, China
    2.Wuhan Optics Valley Information Technology Co., Ltd., Wuhan 510220, China
  • Online:2024-04-15 Published:2024-04-15

摘要: 对于配电线网施工作业这一类高危行业而言,在施工过程中按照安全守则佩戴安全帽是避免事故发生的有效途径之一。由于配电线网施工环境复杂多变,导致现有的安全帽识别方法在自然场景下常出现误检漏检的问题且不能满足实时性需求。为提高自然场景下的安全帽识别准确率以及识别效率,提出一种面向自然场景下配电线网施工的安全帽佩戴识别检测网络模型YOLO-ACON-Attention。该方法以YOLOv5算法为基础,采用自适应判断激活函数取代原有的激活函数,加强模型检测能力。在骨干网络中使用二轮四向IRNN网络构造自适应注意力模块提升模型的图像信息特征提取能力。实验结果证明,与原YOLOv5算法相比,该算法的精确率和召回率为94.75%和89.29%,分别提高了7.65%和5.17%,检测速度为36.5 FPS。

关键词: 安全帽检测, 目标检测, 激活函数, 注意力网络

Abstract: For high-risk industries such as distribution network construction operations, wearing safety helmets in accordance with safety codes during construction is one of the effective ways to avoid accidents. Due to the complex and changeable construction environment of distribution network, the existing safety helmet identification methods often have the problem of false detection and leakage in natural scenarios and cannot meet the real-time requirements. In order to improve the recognition accuracy and efficiency of safety helmets in natural scenes, a safety helmet wearing recognition detection network model YOLO-ACON-Attention for distribution network construction in natural scenes is proposed. Based on the YOLOv5 algorithm, the adaptive judgment activation function is used to replace the original activation function to strengthen the model detection ability. Secondly, the adaptive attention module is constructed by using the two-round and four-way IRNN network in the backbone network to improve the image information feature extraction ability of the model. Experimental results show that compared with the original YOLOv5 algorithm, the accuracy and recall of the algorithm are 94.75% and 89.29%, which are improved by 7.65% and 5.17%, respectively, and the detection speed is 36.5 FPS.

Key words: helmet detection, object detection, activation function, attention networks