Development of Yoga Pose Estimation Model for Smart Healthcare Using Deep Learning

Abstract

Yoga, a discipline that originated in India, is an ancient science that focuses on the body, mind, and soul in a synergistic way. Today, many people consider it an essential part of healthcare and a way of life. Yoga asana (pose) is one of the eight limbs of yoga and, like any exercise, it is crucial to practice it accurately to avoid injury. With the emergence of smart healthcare and the increasing demand for personalized health care, the development of intelligent systems for yoga training and monitoring has become a priority. newlineTo aid individuals in their self-practice of yoga, this research proposed a hybrid convolutional neural network (CNN) and a Gated Recurrent Unit (GRU) deep learning-based yoga pose estimation model that can recognize yoga poses named as quotiSmartYogquot. This real time model is also incorporated feedback mechanism provides users with real-time correction feedback on their postures, enabling them to correct errors and improve their practice. newlineYoga pose is also an action, as involves performing a series of complex movements that begin from a neutral position, progress through a set of intermediate steps, and culminate in a final pose. After holding the pose for a few seconds, the practitioner then returns to the starting position. In this work we have considered yoga pose as an action, so needed a video dataset of yoga asana for model training. But very few datasets are available, so due to the lack of availability of video dataset, so first we have created a large and diverse Yoga pose video dataset of 4 classes i.e. standing poses, sitting poses, prone poses and supine poses, total 24 poses, which consist of 2700 videos created with the help of 31participations (21 female and 10 male ), all poses was performed under the supervision of a yoga trainer. To make the dataset more complex the video is recorded from four directions i.e. front, back, left and right in order to examine yoga pose recognition from four different angles. We named this dataset quotSmartYog Datasetquot and will make it availabl