YoPose: Yoga Posture Recognition Using Deep Pose Estimation
Miniar Ben Gamra, Moulay A. Akhloufi · 2022
Originated in India, yoga is considered a spiritual practice as it brings flexibility, balance, and harmony to both physical and mental health. It becomes the art of healthy living. A variety of positions (also known as asanas) are exercised. Each of them is designed to provide a particular benefit to the body. In contrast, any incorrect action during a yoga session can be harmful to muscles and ligaments. As people are more comfortable with the home workout, the need for an instructor to assess the accuracy of a movement or posture is turned into a need for an auto-guiding framework. Human pose estimation is an important field of research in computer vision. It serves several applications, extending from health monitoring to public safety. As it tackles multiple challenges related to the human posture, it can be used to identify yoga asanas. In this study, we develop a deep-learning self-instruction yoga classifier named YoPose. Based on the pose information, the proposed framework helps individuals to improve their yoga postures by providing personalized feedback. A public dataset including six asanas is used to train and evaluate the model. The introduction of transfer learning and a data augmentation scheme helped to achieve promising results with more than a 98% accuracy.