A novel and intelligent vision-based tutor for Yogāsana: e-YogaGuru

Abstract

Recent days have stamped enormous upsurge about health awareness in society. Self tutoring systems for supervising the performed exercises offer numerous advantages and are therefore emerging as an entity of dire necessity in health-sector. Considering the significantly increasing global acceptance of ‘${{Yog\bar{a}sana}}$’ as one of the most preferred exercise, this paper proposes a novel and an intelligent vision-based self-tutoring system for ${{Yog\bar{a}sana}}$. The proposed system, ’e-YogaGuru’ analyzes the body movements while performing ${{Yog\bar{a}sana}}$, provides feedback about its correctness and further, suggests amendment, if required. Incorporation of angle features in the novel state transition-based approach addresses the earlier reported issues raised due to human anthropometry and variance in the execution speed. Consideration of hold time and suggestion of amendment at two levels, abstract level and detailed amendment (sequences of pre-posture, main-posture and post-posture), make the proposed e-YogaGuru unique and efficient. System is trained for 21 postures derived from the skeleton stream of 8 experts exhibiting variations in anthropometry and execution speed (Knowledge base). A dataset composed of 1750 video sequences (7 ${{Yog\bar{a}sana}}$ performed by 25 practitioners) is used to validate the efficacy of the devised approach. The proposed e-YogaGuru achieved 98.29 % accuracy in correctly identifying the ${{Yog\bar{a}sana}}$ and has been able to suggest required amendment in the incorrectly performed ${{Yog\bar{a}sana}}$ with an accuracy of 96.34 %. Proposed ‘e-YogaGuru’ incorporates significant parameters (hold time and amendment) and achieves appreciable accuracies, thus it not only out-performs the earlier reported systems but also marks a long bounce towards practical deployment.