Yog-Guru: Real-Time Yoga Pose Correction System Using Deep Learning Methods

Self-learning is an integral part in Yoga, but incorrect posture while performing yoga can lead to serious harm to muscles and ligaments of the body. Thus to prevent this we present an intuitive approach based on machine learning techniques to correct the practitioner’s pose while performing various yoga asanas. The proposed system is aimed at providing concise feedback to the practitioner so they are able perform yoga poses correctly and assist them in identifying the incorrect poses and suggest a proper feedback for improvement in order to prevent injuries as well as increase their knowledge of a particular yoga pose. A data-set of Five Yoga pose (i.e. Natarajasana and Trikonasana, and Vrikshasana and Virbhadrasana 1 & 2 and Utkatasana) has been created from collecting images from the Internet as well as from different individuals that took part in development of this system.A deep learning model is proposed which uses convolutional neural networks (CNN) for yoga pose identification along with a human joints localization model followed by a process for identification of errors in the pose for developing the system. Using the proposed system we have been able to achieve a classification accuracy of 95% for pose identification. After obtaining all the information about the pose of the user the system gives feedback to improve or correct the posture of the user.