Yoga Posture Estimation and Correction using Mediapipe and Deep Learning Models

Today’s advancements in computer vision aim to automatically recognize human actions from videos or images and furnish reports on improving posture. Human action surveillance has applications in fields like health care, security, sports, etc. One of the best examples is yoga. Yoga is currently the most effective and convenient form of physical activity for a healthy body operation. But yoga requires a trainer who can constantly monitor the perfectness of various positions executed for perfection and prevent injuries. However, constant human monitoring is not always available. So in this paper, we proposed a simple solution of a system that can track the movement of different body parts and measure the correctness of various yoga postures for the user. The proposed model will also provide correction feedback for yoga poses on pre-recorded videos and in real-time with an accuracy of 93.43%. The paper also discusses previous pose estimation techniques in detail and explains different deep learning models used for yoga pose classification. The past few years have shown major improvements in pose estimation models, particularly in transitioning from single-person to multi-person pose detection in a single frame. These resulted in the creation of many AI-powered training applications.