Computer Vision-Based Real-Time Exercise Pose Detection and Validation: A Web-Based Approach for AI-Assisted Fitness Applications

Abstract

Modern sedentary lifestyles and incorrect exercising methods can have a negative effect on your physical wellbeing, often causing posture related pain and musculoskeletal issues. Exercising regularly and practicing yoga can improve your physical wellbeing. However, it is difficult to maintain correct posture without constant expert oversight. To overcome this issue, we have developed a web based computer vision application called PosePerfect which can detect and validate poses during exercising and yoga in real time. PosePerfect utilizes a client-server architecture model. The client-side is based on React while the server side utilizes Node.js and Express to store data in MongoDB. The PoseNet system identifies and updates the location of 17 body points 25-60 times/second.PosePerfect's rule-based engine functions on the basis of given data including joint angle measurements, relative limb length, and spatial position information. In addition, an app timer is implemented such that a pose is registered by the timer only after 3-5 seconds of sustained accuracy greater than 75% is achieved and/or sustained for that time period as defined by the literature review.
TensorFlow.js is used for all local computation on the user’s device and only summary scores assessing positioning are sent to the cloud. Initial trials have shown a strong correlation (Pearson r=0.87, p<0.001) between evaluation of five different yoga poses utilizing the PosePerfect system and the system has resulted in a mean absolute error of 8.2 units on a 0-10,000 unit scale. Additionally, the system allows for performance to be transferred between devices in real-time and results in a 87% user satisfaction rate and a 93% accuracy rate. his shows the potential benefits of using private computer vision to help with exercise monitoring and prevention of physical health issues using AI.