Scholarlink

<p>Yoga, an ancient science promoting mental and physical balance through asana, meditation, and breathing, is increasingly popular due to modern stress. Accurate yoga pose (asana) detection is vital for self-instructed home practice and applications like human–computer interaction and healthcare. This necessitates robust yet efficient classification models. Therefore in this direction, this study presents an efficient YogCNN framework for an accurate yoga pose classification. The proposed framework leverages the fusion of MobileNetV3-Small for low-latency inference with EfficientNet-B0 for robust multi-scale feature extraction by incorporating squeeze and excitation (SE) to refine the features of both networks in the first stage. Further, bi-directional cross attention is applied to exchange the information between the two refined tokenized feature maps. The architecture sustains lightweight by adopting structured pruning and quantization, reducing both computational cost and model footprints. These optimizations make YogCNN highly practical and deployable solution for mobile pose applications. The proposed framework has been trained and tested on Yoga-82, Yoga-107, and Yoga pose datasets and achieved state-of-the-art performance with computational efficiency. The classification accuracies up to 98.94% of top-1 accuracy and 99.99% of top-5 accuracy for Yoga-82, 97.88% for Yoga-107, and 95.08% for Yoga pose dataset have been observed. An ablation study is carried out to support the experiments. Further, comparative analysis confirms the superiority of proposed framework.</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

YogCNN: an ensemble neural network architecture for yoga asana prediction

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Monika Singh,
  • S. Indu,
  • N. Jayanthi

摘要

Yoga, an ancient science promoting mental and physical balance through asana, meditation, and breathing, is increasingly popular due to modern stress. Accurate yoga pose (asana) detection is vital for self-instructed home practice and applications like human–computer interaction and healthcare. This necessitates robust yet efficient classification models. Therefore in this direction, this study presents an efficient YogCNN framework for an accurate yoga pose classification. The proposed framework leverages the fusion of MobileNetV3-Small for low-latency inference with EfficientNet-B0 for robust multi-scale feature extraction by incorporating squeeze and excitation (SE) to refine the features of both networks in the first stage. Further, bi-directional cross attention is applied to exchange the information between the two refined tokenized feature maps. The architecture sustains lightweight by adopting structured pruning and quantization, reducing both computational cost and model footprints. These optimizations make YogCNN highly practical and deployable solution for mobile pose applications. The proposed framework has been trained and tested on Yoga-82, Yoga-107, and Yoga pose datasets and achieved state-of-the-art performance with computational efficiency. The classification accuracies up to 98.94% of top-1 accuracy and 99.99% of top-5 accuracy for Yoga-82, 97.88% for Yoga-107, and 95.08% for Yoga pose dataset have been observed. An ablation study is carried out to support the experiments. Further, comparative analysis confirms the superiority of proposed framework.