Yoga Pose Recognition Based On Convolutional Neural Networks

People are more accepting of virtual yoga courses now than they were before the epidemic, and this preference for online forms is likely to continue even after the outbreak is over. Thus, an automated yoga self-training system is significant. Automatic yoga position recognition is necessary for the practitioner to carry out their yoga exercise with accurate poses at any time. Convolutional neural networks (CNN) is a deep learning architecture that has been proven to be a powerful tool in image classification. This dataset was subjected to an experiment using many common CNN architectures. The experiment, however, excludes the use of certain powerful image classification models such as VGG-16, AlexNet, and GoogleNet. Therefore, a study of classifying 82 yoga classes using the VGG16, AlexNet, and GoogleNet models via transfer learning is carried out to evaluate their performance in terms of sensitivity, precision, F1-score, specificity, Top-1 accuracy, and Top-5 accuracy. Transfer learning is a handy technique for eliminating time-consuming computations and allowing image classification with a smaller dataset. GoogleNet was the most appropriate model for classifying 82 yoga classes since it achieved the highest percentages of 99.94%, 80.14%, 92.97%, 85.93%, 87.89%, and 97.08% for specificity, sensitivity, precision, F1-score, Top-1 accuracy, and Top-5 accuracy, respectively.