Real-time exercise posture classification using LSTM and pose estimation in unconstrained gym environments

Maintaining correct posture during strength training is essential for optimizing performance and preventing injuries. However, automated posture evaluation systems often face challenges in real-world gym environments due to variations in lighting, clothing, and body types. This study presents a real-time posture classification system that integrates pose estimation with deep learning to evaluate squats, overhead presses, and triceps pushdowns. A custom annotated dataset was created from scratch by recording 3- to 4-s video clips of individuals performing the target exercises under varied lighting, backgrounds, and user conditions in natural gym settings. Pose landmarks were extracted using MediaPipe, and temporal patterns were modeled with an LSTM network trained on sequences of 30 frames containing 99 pose features (33 landmarks × 3 coordinates ( x , y , z )) per repetition. The model achieved test accuracies of 90% for squats and overhead presses, and 95% for triceps pushdowns. The system was deployed using a standard webcam and demonstrated responsive real-time feedback without requiring specialized hardware, making it practical for live posture monitoring. This lightweight and accessible approach demonstrates the potential of pose estimation and deep learning for practical real-time posture assessment, with applications in personal training, injury prevention, and rehabilitation.