Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (24): 313-320.DOI: 10.3778/j.issn.1002-8331.2107-0005

• Engineering and Applications • Previous Articles    

Improved Detection Algorithm and Its Application in Safety Control in Substation Scenario

WU Hongyi, LEI Jingsheng, CHEN Linfeng, YANG Shengying   

  1. School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Hangzhou 310000, China
  • Online:2022-12-15 Published:2022-12-15

改进目标检测算法在变电站内安全管控的应用

吴宏毅,雷景生,陈林锋,杨胜英   

  1. 浙江科技学院 信息与电子工程学院,杭州 310000

Abstract: As an important part of the power system, it is crucial to ensure the safety of substation operators. In order to automatically detect whether the operators are correctly wearing overalls and helmets, a Transformer-based self-attentive coding feature fusion lightweight target detection network is proposed. The attention mechanism fuses multi-scale features by using a lightweight backbone network to extract features. A quality focus loss method is proposed to improve the inconsistent inference process in the training and testing phases of the target detection model. Meanwhile, 5,200 pieces of personnel workwear helmet data are collected and labeled in the substation scene. The proposed lightweight target model is trained on the homemade workwear helmet dataset and validated on the test set, and the target detection method recognizes mAP up to 44.6% and AP50 up to 79.5%, reaching 117 FPS.

Key words: object detection, deep learning, light weight, Transformer, neural network

摘要: 变电站作为电力系统中重要一环,保证变电站作业人员的安全是至关重要的。为了自动检测作业人员是否正确穿着工作服佩戴安全帽,提出一种基于Transformer自注意力编码特征融合轻量级的目标检测网络。通过采用轻量级的主干网络提取特征,注意力机制融合多尺度特征。提出了质量焦点损失方法,改善目标检测模型训练和测试阶段推理过程不一致问题。同时,采集并标注变电站场景下人员工作服安全帽数据5?200张。将提出的轻量级目标模型在自制的工作服安全帽数据集上训练,并在测试集上验证,该目标检测方法识别mAP达44.6%,AP50达79.5%,达到117 FPS。

关键词: 目标检测, 深度学习, 轻量化, Transformer, 神经网络