Scholarlink

Deep learning in HAR has opened numerous opportunities for advanced fitness monitoring and healthcare applications, making it a prominent area of study in computer vision. Since the pandemic, yoga has gained worldwide attention due to rising stress levels and workload in modern life. There are innumerable resources for learning yoga, and many people prefer self-learning, which increases the risk of incorrect posture practice, potentially leading to physical injuries and long-term health issues. We propose a deep learning-based yoga pose classification system that achieves good results using limited and publicly available data. We examine spatial-temporal feature extraction for MoveNet multi-output heatmaps and regression, evaluated on the Yoga Poses Dataset which consists of five yoga poses, including Tree Pose and Warrior and Triangle’s Posture. Combined use of CNN-LSTM architecture enables our model to reach 95.74% accuracy when operated with batch size 16 and it surpasses individual CNN performance of 94.39% and LSTM performance at 93.27% because it effectively captures sequential posture transitions alongside anatomical keypoint dynamics. The system operates effectively on the NVIDIA RTX 3090 GPU alongside JupyterLab 4.0.9 to demonstrate real-time suitability for personal fitness monitoring applications through its strong assessment metrics of 95.33% F1-score and 99.32% AUC-ROC. This research creates a connection between motion detection and activity recognition, which delivers an expandable system to use artificial intelligence for yoga posture correction and safety enhancement.

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

Yoga Pose Recognition Using MoveNet and Hybrid CNN-LSTM: A Deep Learning Approach for Posture Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Diya Koyani,
  • Devansh Javia,
  • Samarth Dhol,
  • Nishant Kanani,
  • Hemang Thakar,
  • Amit Thakkar

摘要

Deep learning in HAR has opened numerous opportunities for advanced fitness monitoring and healthcare applications, making it a prominent area of study in computer vision. Since the pandemic, yoga has gained worldwide attention due to rising stress levels and workload in modern life. There are innumerable resources for learning yoga, and many people prefer self-learning, which increases the risk of incorrect posture practice, potentially leading to physical injuries and long-term health issues. We propose a deep learning-based yoga pose classification system that achieves good results using limited and publicly available data. We examine spatial-temporal feature extraction for MoveNet multi-output heatmaps and regression, evaluated on the Yoga Poses Dataset which consists of five yoga poses, including Tree Pose and Warrior and Triangle’s Posture. Combined use of CNN-LSTM architecture enables our model to reach 95.74% accuracy when operated with batch size 16 and it surpasses individual CNN performance of 94.39% and LSTM performance at 93.27% because it effectively captures sequential posture transitions alongside anatomical keypoint dynamics. The system operates effectively on the NVIDIA RTX 3090 GPU alongside JupyterLab 4.0.9 to demonstrate real-time suitability for personal fitness monitoring applications through its strong assessment metrics of 95.33% F1-score and 99.32% AUC-ROC. This research creates a connection between motion detection and activity recognition, which delivers an expandable system to use artificial intelligence for yoga posture correction and safety enhancement.