Scholarlink

<p>Human‒computer interaction (HCI) plays a crucial role in augmented reality (AR). This paper introduces SFG-YOLOv8, an efficient and lightweight gesture keypoint detector designed for low-computational-power AR scenarios. We propose SFG-ShuffleNetV2 as the backbone network by improving the convolution operation, resulting in fewer parameters, and introduce the contextual bottleneck refinement module (CBRM) to enlarge the receptive field, enhancing the capture of the global context. A small feature keypoint detection layer (SFKDL) is incorporated into the feature extraction pipeline to improve the detection accuracy of small gesture features. Additionally, the enhanced coordinate attention module (E-CAM) is employed to address the limitations of single-scale feature representations, improving the ability to capture multiscale information. The experimental results on our AR-HGRI dataset demonstrate that SFG-YOLOv8-N and SFG-YOLOv8-S achieve mean average precision (mAP) scores of 71.2% and 73.9%, with inference FPSs of 119.4 and 94.1, respectively. These improvements—7.4% and 5.6% greater mAPs than those of YOLOv8-N and YOLOv8-S—highlight the effectiveness of SFG-YOLOv8 for gesture keypoint detection in augmented reality environments.</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

SFG-YOLOv8: efficient and lightweight small-feature gesture keypoint detector

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月20日

查看全文
  • Weimin Che,
  • Hui Zhang,
  • Bangxue Wu,
  • Qun Li,
  • Hongji Zhang,
  • Shilong Yuan,
  • Hongcheng Yang

摘要

Human‒computer interaction (HCI) plays a crucial role in augmented reality (AR). This paper introduces SFG-YOLOv8, an efficient and lightweight gesture keypoint detector designed for low-computational-power AR scenarios. We propose SFG-ShuffleNetV2 as the backbone network by improving the convolution operation, resulting in fewer parameters, and introduce the contextual bottleneck refinement module (CBRM) to enlarge the receptive field, enhancing the capture of the global context. A small feature keypoint detection layer (SFKDL) is incorporated into the feature extraction pipeline to improve the detection accuracy of small gesture features. Additionally, the enhanced coordinate attention module (E-CAM) is employed to address the limitations of single-scale feature representations, improving the ability to capture multiscale information. The experimental results on our AR-HGRI dataset demonstrate that SFG-YOLOv8-N and SFG-YOLOv8-S achieve mean average precision (mAP) scores of 71.2% and 73.9%, with inference FPSs of 119.4 and 94.1, respectively. These improvements—7.4% and 5.6% greater mAPs than those of YOLOv8-N and YOLOv8-S—highlight the effectiveness of SFG-YOLOv8 for gesture keypoint detection in augmented reality environments.