A novel machine learning-based shoulder exercise assistant program

As health awareness increases in modern society, interest in exercise is surging. In particular, as home training has become popular due to the COVID-19 pandemic, and shoulder injuries have become more frequent due to the increase in the aging population in Korea, there is a growing need for a system that can perform accurate exercises at home. We propose a real-time motion classification and feedback system that allows shoulder exercises to be monitored and evaluated at home. We implement a system that classifies shoulder exercise movements using annotated keypoint data (wrist, elbow, and shoulder) obtained through CVAT, and checks exercise accuracy in real time by training yolov5’s object detection model. The system utilizes six machine learning models: Random Forest Classifier, Logistic Regression Classifier, Gradient Boosting Classifier, Histogram-based Gradient Boosting Classifier, SVM Classifier, and Extra Trees Classifier, and utilizes cross-validation and sampling techniques. This enables the identification of critical features that contribute to successful exercise performance. The system also provides real-time feedback to the user to improve their behavior during exercise. This approach demonstrates the applicability of human posture estimation and motion classification in real-world settings and highlights the importance of keypoints in shoulder exercise performance for accurate feedback and instruction.