LiftMate: Gym Exercise Posture Correction and Classification

Correct exercise form is essential to prevent injuries and optimize performance. This paper introduces LiftMate, a computer vision based system for gym exercise posture classification in biceps curl, plank, and squat using machine learning and deep learning models. Our dataset consists of self-recorded and public videos, with MediaPipe employed for real-time keypoint detection. We extract spatial features (joint angles and distances) and evaluate multiple classification models. Experimental results demonstrate that deep models achieve near-perfect accuracy. In particular, our 3-layer neural network attained almost 100% classification accuracy on all three exercises, significantly outperforming traditional classifiers (e.g. logistic regression). These findings confirm the potential for fully automated real-time posture assessment in fitness. While our models achieved very high accuracy, limitations such as dataset diversity and size suggest avenues for future work.