Scholarlink

<p>Keypoint detection is a core task in the visual domain, facing challenges such as occlusion, false detection, and complex poses. To address these issues, an efficient keypoint detection algorithm, EKD-YOLO, based on the YOLOv8s model, is proposed. Firstly, in case of enhancing the model’s feature fusion capability, the Fusion-Neck architecture is depicted to replace the original Neck part of YOLOv8. Secondly, for the purpose of improving model efficiency, the feature extraction module has been optimized by proposing a more lightweight C2fiAFF module and introduces the SCDown module. Furthermore, in order to improve accuracy and precision, an efficient attention mechanism, C2iEMA is designed. Experimental data indicate, compared to the YOLOv8s model, the improved EKD-YOLO algorithm achieved a 2.5% increase in mAP@0.5, a 4% improvement in P%, and a reduction in FLOPs and parameter count. The overall performance of the model was enhanced. Our code and models are available at <a href="https://github.com/geleerde/EKD-YOLO">https://github.com/geleerde/EKD-YOLO</a>.</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

An efficient human pose keypoint detection algorithm base on YOLOv8

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Shengqi Yang,
  • Haiying Liu,
  • Yifeng Du,
  • Tao Sun,
  • Chaoqun Wang,
  • Fei Lin,
  • Lida Liu

摘要

Keypoint detection is a core task in the visual domain, facing challenges such as occlusion, false detection, and complex poses. To address these issues, an efficient keypoint detection algorithm, EKD-YOLO, based on the YOLOv8s model, is proposed. Firstly, in case of enhancing the model’s feature fusion capability, the Fusion-Neck architecture is depicted to replace the original Neck part of YOLOv8. Secondly, for the purpose of improving model efficiency, the feature extraction module has been optimized by proposing a more lightweight C2fiAFF module and introduces the SCDown module. Furthermore, in order to improve accuracy and precision, an efficient attention mechanism, C2iEMA is designed. Experimental data indicate, compared to the YOLOv8s model, the improved EKD-YOLO algorithm achieved a 2.5% increase in mAP@0.5, a 4% improvement in P%, and a reduction in FLOPs and parameter count. The overall performance of the model was enhanced. Our code and models are available at https://github.com/geleerde/EKD-YOLO.