Real-Time Posture Correction of Squat Exercise: A Deep Learning Approach for Performance Analysis and Error Correction

The squat is a widely performed exercise essential for maintaining physical fitness and overall well-being. It involves a series of movements in which an individual lowers their torso by bending the knees, and then returns to the starting position. Incorrect squat execution can lead to physical injuries, highlighting the importance of expert guidance. However, many individuals cannot afford professional training due to high costs, necessitating an automated workout analysis system. Existing vision-based squat analysis models face challenges related to occlusion and environmental factors. To address these limitations, we propose a method that utilizes squat videos captured by a stereo camera. The MediaPipe algorithm extracts and tracks keypoints from video frames. Detected keypoints are then used to classify squat types and evaluate squat performance using a Bi-CGRU model. Additionally, we introduce a regression-based approach to predict the correct squat posture. A deep learning-based regressor, trained on a dataset of squat videos, estimates the correct joint angles. When an improper squat is detected, the system provides feedback by presenting the correct squat form to the participant. Experimental results demonstrate that the Bi-CGRU classification model achieves 96.1% accuracy, while the pose correction regression model attains 99.0% accuracy using mean squared error loss. Both qualitative and quantitative evaluations indicate that our approach outperforms state-of-the-art methods utilizing stereo video input. The system helps users to refine their squat technique and reduce the risk of injury by offering real-time posture correction based on biomechanics guidance.