Yoga Pose Recognition and Correction Using Deep Learning
C. S. Kanimozhi Selvi, K.S. Kalaivani, S. Srigha, Hariesh Ramesh, S.L. Kalki Kartik · 2024
Yoga is an ancient practice that holds contemporary significance in promoting holistic well- being. This study focuses on yoga pose recognition via deep learning-based methodology, utilizing Media Pipe for pose estimation. Specifically, the study concentrates on estimating poses for beginners. The yoga pose images were manually categorized into three levels: Beginners, Intermediate, and Advanced. The estimated yoga pose data for beginners was used to train several types of deep learning models: an Advanced Convolutional Neural Network (CNN), a hybrid model combining Convolutional Neural Network with Long Short-Term Memory, ResNet50, VGG16. All the models have demonstrated promising accuracy. The CNN model achieved a training accuracy of $\mathbf{9 8. 8 9 \%}$ and a validation accuracy of $83.33 \%$. In contrast, the CNN- LSTM model achieved a training accuracy of $94.06 \%$, resulting in a marginal increase in validation accuracy to $84.79 \%$. Furthermore, with an accuracy of $92 \%$, the VGG16 model showed excellent performance, while the ResNet50 model attained an astounding $\mathbf{9 3. 6 8 \%}$ accuracy. The primary objective is for the model to accurately identify beginners’ yoga poses and provide recommendations for adjusting alignment to achieve the perfect Yoga Pose.