YogaHelp: Leveraging Motion Sensors for Learning Correct Execution of Yoga With Feedback

Abstract

By leveraging motion sensors, the study of physical activities has become effortless and convenient. Yoga is an excellent form of physical activity or exercise that involves all parts of the body. Regular practice of yoga improves both the physical and mental health of the people, if the practitioner performs it correctly. Though the existing exercise recognition methods have shown quite a good performance using deep neural networks, they fail to assess the correctness of execution. Feedback describing correctness level is very useful to an amateur practitioner. In this article, we aim to develop a YogaHelp system to help amateurs for learning the correct execution of yoga without any supervision of a trainer. The importance of such a system is much higher during a pandemic situation, where the trainers can not be availed due to the risk of catching an infection. YogaHelp leverages the motion sensors including accelerometer and gyroscope, to recognize 12 linked-steps of sun salutation (Surya Namaskar) yoga along with their correctness level. YogaHelp employs a deep learning model built using convolutional layers for yoga step recognition without explicit feature extraction. The novelty of the system lies in the feedback that describes the speed of execution and angular deviation of the posture from the standard ones. We develop a prototype for collecting a training dataset from eight professional yoga trainers and for validating the effectiveness of the system.