Scholarlink

<p>Novel human object posture detection method for volleyball videos based on YOLO (You Only Look Once) &amp; high-resolution deep learning is put forth. The network of YOLO is realized by lightweighted GhostNet network and CBAM (Convolutional Block Attention Module). The high-resolution deep learning is designed for light-weighting, which significantly improves the performance and efficiency of the human object posture estimation model and enables multi-person posture estimation. The neck is used as the central point to establish a coordinate system for preprocessing, and proportions are adjusted to quantify numerical values such as action angles and speeds, which are then utilized to evaluate the quality of players’ actions. The Dynamic Time Warping (DTW) algorithm is employed to compare players’ actions with preset actions and analyze the results of action quality. Experimental results on the Max Planck Institute for Informatics (MPII) and Common Objects in Context (COCO) datasets, as well as genuine volleyball match videos, indicate that the proposed algorithm retains a high degree of human posture estimation accuracy while substantially reducing the number of parameters and computational demands compared to the original model. Our method not only is suitable for volleyball videos as the main application scenario, but also it can be extended to other sports or to multi-object dense scenes.</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

Novel approach of human object posture detection for volleyball videos based on YOLO & high-resolution deep learning

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Degan Zhang,
  • Xuejie Ren,
  • Ting Zhang,
  • Chuanpeng Bao,
  • Chenhui Dou,
  • Tao Liu

摘要

Novel human object posture detection method for volleyball videos based on YOLO (You Only Look Once) & high-resolution deep learning is put forth. The network of YOLO is realized by lightweighted GhostNet network and CBAM (Convolutional Block Attention Module). The high-resolution deep learning is designed for light-weighting, which significantly improves the performance and efficiency of the human object posture estimation model and enables multi-person posture estimation. The neck is used as the central point to establish a coordinate system for preprocessing, and proportions are adjusted to quantify numerical values such as action angles and speeds, which are then utilized to evaluate the quality of players’ actions. The Dynamic Time Warping (DTW) algorithm is employed to compare players’ actions with preset actions and analyze the results of action quality. Experimental results on the Max Planck Institute for Informatics (MPII) and Common Objects in Context (COCO) datasets, as well as genuine volleyball match videos, indicate that the proposed algorithm retains a high degree of human posture estimation accuracy while substantially reducing the number of parameters and computational demands compared to the original model. Our method not only is suitable for volleyball videos as the main application scenario, but also it can be extended to other sports or to multi-object dense scenes.