Development of Machine Learning Based Real-Time Squat Training Feedback System

Regular exercise is crucial for maintaining good health, as it promotes muscle growth and helps prevent cardiovascular diseases. Among various forms of exercise, multi-joint exercises are considered the most effective for individuals with limited time availability. However, unsupervised multi-joint exercises may be ineffective and can even lead to injuries. Hence, technological intervention during the workout is required to improve the quality and safety of the training when supervisors are unavailable. Therefore, an automatic recording system for squats with prompt feedback is proposed in this study. Users could analyze their movements using this system and receive suggestions through the screen to improve their form and perform squats correctly even when the coach is not around. To provide feedback immediately, the input features of the machine learning model had to be simple and accurate. Hence, instead of using the entire video, only three critical features were selected in this study to train the machine learning model. The first feature was the angle of the body and thigh (BT), and the second feature was the backward bending of the foot (Dorsiflexion, DF). The third feature was bar-shift (BS), which is the deviation between the barbell and virtual center line (extending from the middle of the ankle and forefoot). In this study, 1826 squats from 54 subjects were successfully recorded and labeled to 11 different conditions. The recorded features were processed to create six datasets. Then, five machine-learning architectures, including Random Forest, XGBoost, 1D-CNN, LSTM, and LSTNet, were trained on different combinations of datasets to find the optimized model. Among them, Random Forest showed the best accuracy in predicting the quality of the squat (72.6%) and recognizing the functional disabilities that led to poor squatting. Finally, a real-time squat training feedback system was demonstrated and examined. Three trainers with an advanced barbell squat technique were asked to perform 10 good squats and 10 questionable squats. The proposed system successfully recorded 54 out of 60 squats, and the accuracy of rating the squat was 55.5%.