Scholarlink

<p>Accurate identification and fine analysis of sports actions in real-time videos for athletes, especially fitness exercises, is a difficult task in computer vision and remains a challenge due to the similarity of some actions. Aiming to improve the athletes’ performances and the robustness of actions classifying systems from videos sport in both real-time and pre-recorded scenarios, we propose in this paper a deep learning model for classifying 16 fitness exercises based on normalized key body landmarks. The MediaPipe pose estimation model is used to extract the key body landmarks. These latter are then used as training data for the learning model to accurately predict exercises poses. Finally, two measures are defined using the extracted key landmarks to provide a detailed and comprehensive analysis of the exercises. Various experiments are conducted on real and pre-recorded videos. The obtained results demonstrate the performance of our approach. The proposed framework gives promising results in effectively classifying fitness exercises, offering potential advancements in monitoring and guidance within fitness training programs in order to evaluate and improve the performance of athletic actions.</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

Efficient framework for sport actions identification and analysis in real-time fitness videos using neural network

  • 期刊论文

  • 发表日期：

    2025年10月4日

查看全文
  • Nacéra Laiche,
  • Boutiche Imad Ahmed,
  • Diab Mohamed Hicham

摘要

Accurate identification and fine analysis of sports actions in real-time videos for athletes, especially fitness exercises, is a difficult task in computer vision and remains a challenge due to the similarity of some actions. Aiming to improve the athletes’ performances and the robustness of actions classifying systems from videos sport in both real-time and pre-recorded scenarios, we propose in this paper a deep learning model for classifying 16 fitness exercises based on normalized key body landmarks. The MediaPipe pose estimation model is used to extract the key body landmarks. These latter are then used as training data for the learning model to accurately predict exercises poses. Finally, two measures are defined using the extracted key landmarks to provide a detailed and comprehensive analysis of the exercises. Various experiments are conducted on real and pre-recorded videos. The obtained results demonstrate the performance of our approach. The proposed framework gives promising results in effectively classifying fitness exercises, offering potential advancements in monitoring and guidance within fitness training programs in order to evaluate and improve the performance of athletic actions.