DEEP LEARNING BASED YOGA LEARNING APPLICATION

Abstract

Yoga has been into practice for thousands of years and in today’s era it is highly suggested for people to practice it to revitalize them in their physical and mental wellness. Although yoga is considered safe, there are several asanas that need to be practiced under the guidance and supervision of a trained instructor. Using human action recognition techniques for yoga pose estimation, we have tested and designed an automated yoga asana identification and correction system that performs well in the absence of a trained instructor. This application finds its importance specially the times of the COVID-19 epidemic where social distancing is the new norm.

The application identifies yoga asana being practised by the user in real time with the help of deep learning techniques based on convolutional neural networks (CNN) and transfer learning which might require correction. The application corrects the user for right yoga posture using multi-person 2D pose estimation algorithm called OpenPose. It uses multiple deep learning algorithms for pose estimation.

The application is designed for 18 different asanas which the users can practice for a start. The system can predict an asana with over 87.6% accuracy considering the fact that user can face the camera with multiple different views while practicing yoga, namely left-hand side view, right-hand side view or front view; depending on the pose. This is taken into consideration so that it gets very much easier for the user to practice yoga without thinking about how the user should be facing the camera hence we can say that it is more user friendly. The system corrects the user by showing the direction in which they should move the body part, which is not in the correct position in real time, as a feedback to the user. As and when the user changes its posture the feedback to the user changes till the user gets the right posture.