Smart Yoga: Machine Learning Approaches for Real-Time Pose Recognition and Feedback

Abstract

Yoga is an ancient science and discipline with a long history associated with India. It helps in maintaining a person's physical fitness as well as providing mental harmony at the same time. Due to the stress levels in modern life, yoga has recently acquired international attention. Although there are many ways to learn yoga and a variety of materials available, doing yoga without proper instruction can result in major problems like acute pain and long-term chronic issues. To overcome all these major issues, we have proposed an app that identifies yoga poses, which are performed by the user and outputs voice feedback, about the current pose of Asan. Since finding the correct relevant dataset was the major issue, we used web scraping to scrap multiple images from the web and made our dataset to train our model. The proposed model in this study is a KNN binary classifier that classifies whether the asan is correct or not, through this it allows real-time pose estimation to detect the error in a person's pose, thereby allowing them to correct it. The proposed model has shown an accuracy of 0. 836 % and an F1-Score of 82.3% for yoga pose detection and estimation.