Yoga Poses Identification and Classification Based on Convolutional Neural Network and Transfer Learning with Media Pipe

Abstract

Yoga is an ancient Indian discipline that promotes mental and physical well-being. It's become popular due to the stress of modern life. There are many ways to learn yoga, including studios, private instructors, and online resources. Many students of yoga struggle to identify their own mistakes when learning on their own. This article proposes a new approach for the effective identification and classification of different yoga poses using deep learning algorithms. The Media-pipe library is used to extract user-relevant features from 85 videos featuring 15 yoga practitioners doing 6 different poses. In the study, results from many deep learning models are compared, both with and without extracting features. Several different learning models achieved their best performance when fed skeletonized pictures to a neural network for training. Results from several models are compared in order to demonstrate the beneficial effect of skeletonization. With a validation accuracy of 99.9% on non-skeletonized images, Mobile-Net with CNN outperforms CNN, LSTM and SVM by a wide margin. Skeletonized images are used by the proposed model MobileNet, which achieves an accuracy result of 99.9%.