Scholarlink

<p>In the object detection tasks, it is challenging to detect small objects in disordered or overlapping regions of images. This paper proposes a Personal Protective Equipment (PPE) detection algorithm based on an improved YOLOv8s model to address these issues. Firstly, an improved Global Attention Mechanism (iGAM) is introduced based on the original Global Attention Mechanism (GAM) to achieve a larger receptive field for aggregating data while reducing computational costs. The iGAM module is integrated into the detection architecture of YOLOv8s. Secondly, the bounding box regression loss with a dynamic focus mechanism decreases the harmful gradient generated by low-quality samples. The proposed model was trained on the Colors Helmet and Vest (CHV) dataset. Experimental results demonstrate that the proposed algorithm achieved a mAP@50 and mAP@50–95 outperformed the original YOLOv8s by 1.3% and 3%, respectively. Furthermore, the algorithm was successfully deployed on the Jetson Orin Nano kit, achieving a processing speed of 30 frames per second (FPS), making it appropriate for real-time applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Personal protective equipment detection using YOLOv8 with improved global attention mechanism and dynamic focus mechanism

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Quang Vinh Truong,
  • Hien Long Pham

摘要

In the object detection tasks, it is challenging to detect small objects in disordered or overlapping regions of images. This paper proposes a Personal Protective Equipment (PPE) detection algorithm based on an improved YOLOv8s model to address these issues. Firstly, an improved Global Attention Mechanism (iGAM) is introduced based on the original Global Attention Mechanism (GAM) to achieve a larger receptive field for aggregating data while reducing computational costs. The iGAM module is integrated into the detection architecture of YOLOv8s. Secondly, the bounding box regression loss with a dynamic focus mechanism decreases the harmful gradient generated by low-quality samples. The proposed model was trained on the Colors Helmet and Vest (CHV) dataset. Experimental results demonstrate that the proposed algorithm achieved a mAP@50 and mAP@50–95 outperformed the original YOLOv8s by 1.3% and 3%, respectively. Furthermore, the algorithm was successfully deployed on the Jetson Orin Nano kit, achieving a processing speed of 30 frames per second (FPS), making it appropriate for real-time applications.