A Deep Learning Framework for Classifying and Evaluating Yoga Exercises

Abstract

Yoga is a type of physical activity that combines physical postures with breathing techniques. Cardiovascular disease is the leading cause of illness and mortality. Yoga's potential involvement in cardiac recovery can be incorporated into cardiovascular rehabilitation. Classifying and evaluating yoga exercises using a deep learning framework plays a significant role in postoperative recovery. Home based rehabilitation often suffers from a lack of patient adherence to prescribed exercise routines, resulting in longer treatment times and higher healthcare expenses. This study proposes a framework for classifying and evaluating the quality of yoga exercises. The ConvNet followed by the deep belief network model is used to create quality scores for input movements to evaluate the quality of yoga asanas. Additionally, various machine learning algorithms are deployed to identify the best performing algorithm. The proposed framework is tested using the COCO dataset and the publicly available dataset from KAGGLE. The model is implemented using a three stage pipeline. In the first stage, an estimator and detector model is used for feature extraction. In the second stage, restricted Boltzmann machines are used for dimensionality reduction in two steps: The forward pass and the backward pass. Finally, the ConvNet followed by the deep belief network is used to classify yoga postures. The quaternion data is treated as a multivariate Gaussian variable to create Markov random fields. To quantify yoga asanas, the Markov random fields network is triggered, and the resulting quaternion point is compared to standard pose data, increasing the likelihood that the yoga position is correct. The importance of our study is that it allows participants to watch their movement in real-time and to obtain a numerical estimate of the accuracy of their yoga stance. Using this method, we were able to achieve a precision of 99.99%, which was previously unattainable on this dataset. Our model is more robust than previous models in terms of performance.