Machine Learning-Based Exercise Posture Recognition System Using MediaPipe Pose Estimation Framework
Weeriya Supanich, Suwanee Kulkarineetham, Parinya Sukphokha, Patcharathon Wisarnsart
Abstract
Daily exercise is essential for good health, but incorrect posture during exercise can lead to pain and injury, especially for the elderly. Hiring a personal trainer can be expensive, and not everyone has access to one. This paper proposes a posture classifier system that uses machine learning to recognize various exercise postures. Instead of having a personal trainer, the goal of this research is to create an automated model that can properly assess exercise posture. From video datasource recorded by a fitness specialist through a simple web camera, we extract body skeleton sequences using the MediaPipe pose estimation framework and evaluate the performance of different machine learning models in detecting each posture class in each type of exercise using precision, recall, and accuracy metrics. The system achieves an average accuracy score of 100% on our test data on three types of exercises, demonstrating its potential as an affordable and accessible solution for monitoring and correcting body postures during exercise.