A Machine Learning Approach Towards Yoga Exercise Classification for Pain Recovery

Introduction

Back injury leading to pain is caused by weak muscles, particularly core and pelvic muscles, by which our regular activities are often hampered (Ghosh et al., 2020). In studies, strengthening activities are shown as effective in alleviating back pain. To improve the condition, specific workouts, and stretches can help the muscles for a stronger lower and upper back. Motion-sensing technology based on depth cameras, like the Kinect sensor, has made it possible to access 3D posture without requiring a complicated setup (Saha & Das, 2021). After getting prescribed specific exercises by a physician, most patients prefer doing them at home due to the lengthy time commitment and high cost of clinical visits (Mousavi & Khademi, 2014). Rehabilitation exercises at home without or with less guidance from therapists can collect real-time assessment, feedback, and supervision through a Kinect-based system (Ghosh & Saha, 2020). Although many such systems have been presented to help increase the likelihood that the patient would perform the recommended exercises correctly, many cannot provide timely feedback (Solongontuya et al., 2021).

In (Bhatlawande et al., 2022), 3D human body joint motion trajectories in Euclidean distances are retrieved as time series from training videos using Microsoft’s Kinect sensor. Then, the time series is expressed as an autoregressive and moving average to model each motion's dynamic process. However, the subjects are asked to perform each movement only once, resulting in a small dataset for training, whereas the proposed work overcomes the limitation by including 5 trials of each exercise performed by each subject, which makes the training data sufficient to avoid overfitting. In (Örücü and Selek 2020), a rule-based expert framework for monitoring and improving training exercises for athletes has been provided. It is simple to dynamically modify the parameters used in the rules, such as the starting and finishing of pose orientation and the tolerance values. The avatar-based guidance system validates the algorithm's performance, however, for optimal performance, the rules must be manually designed. To effectively represent the posture context, the domain of the rules must also be chosen depending on the nature of the postures, which makes it ineffective to expand to a large range of movement, whereas the proposed data-driven approach overcomes the difficulty by evaluating the postures with prior knowledge obtained from a training database.

Researchers frequently have focused on using depth videos rather than the conventional RGB videos because it is simpler to segment the foreground human subject even when the scene is cluttered. Also, since color information is not preserved by depth videos, the segmentation procedure is unaffected by the color of the human subject's clothing (Franco et al., 2020). However, depth data is often vulnerable to noise and does not guarantee satisfactory accuracy, whereas the proposed framework overcomes the issue by using skeleton joint-based data captured by Kinect. It is possible to achieve reliable classification results using supervised machine learning (ML) approaches, like Decision Tree (DT) (Biswas et al., 2020), Random Forest (RF) (Bag et al., 2023), Support Vector Machine (SVM) (Ghosh & Saha, 2020), or deep learning (DL) models like Convolutional Neural Networks (CNN) (Ghosh & Saha, 2022), Recurrent Neural Networks (Ghosh & Saha, 2021), etc. (Ahmed et al., 2015). In (Dajime et al., 2020), the movement competency screen score was calculated using joint kinematics, and logistic regression analysis was used to assess the quality of exercises, but only male subjects were considered for data collection to avoid the effects of gender difference in kinematics. Instead of avoiding it, the proposed model deals with the body structure variability with normalized features.