A Fitness Guider Application for Wheelchair Users Using Machine Learning

Abstract

Exercise is a must for a healthy life. People with disabilities also have to exercise to maintain their healthy life but they can face some injuries and health problems when they train without any guidance. Without a trained eye, it's hard to check whether the exercises are successful or not, and if there are any issues with the posture and the technique of the exercise, it can cause them lots of health problems in the long run. This project focuses on addressing this issue by developing an application that detects the exercise pose of wheelchair users and provides personalized, detailed recommendations on how the user can improve their form. Using human pose estimation with deep learning algorithms to detect users' poses and the system evaluates the vector geometry of the pose through an exercise to provide useful feedback. Based on personal training guidelines, the system uses a dataset of exercise videos of correct and incorrect forms to train the deep learning model. Pose Trainer contains exercises that can be performed by the user without any observation and will be available for Android Users. Initial evaluation of the prototype achieved a 96% accuracy rate, with an AUC-ROC of 0.93 and a confusion matrix revealing a 2% false-positive rate. The model's real-time processing capability allows it to handle up to 1,000 transactions per second.