Advanced Yoga Pose Estimation: Enhancing PoseNet with Adaptive Key Point Elimination
K. Aarthy, A. Alice Nithya, M.N. Kruthi, Roshan Upadhyay, Jahnavi Darbhamulla, Sounak Singh, M. M. Pavikars · 2024
Yoga is an ancient practice famous for its physical and mental benefits and assistance, is gaining recognition as a means of achieving enhanced flexibility, strength, and mental clarity. With the increasing demand for accessible and effective yoga instructions through online mode, in this work a yoga pose estimation model will be developed which may be used with a web application, to pro-vide yoga lessons in the convenience of one's own environment. This paper delves into the integration of advanced computer vision and machine learning techniques, particularly PoseNet, which accurately tracks human poses using 17 key points. But all the 17 key points identified using posenet architecture will not be useful in estimating the pose accuracy and might result in increased redundancy and thus decreasing model performance. To overcome this drawback, an innovative approach called adaptive key point elimination technique is proposed to improve the precision of yoga pose estimation. The proposed methodology involves leveraging transfer learning to fine-tune the PoseNet model for specific yoga poses and employing a custom Convolutional Neural Network (CNN) for the classification of various yoga postures. The experimental methods are tested on a yoga dataset taken from Kaggle website for the following yoga asanas: Chair, Cobra, Dog, Tree, and Warrior poses. The proposed model is able to achieve a validation accuracy of 98.678% and a final evaluation accuracy of 99.32% on the test dataset for our yoga pose estimation system.