AI-Driven Real-Time Exercise Repetition Counting System

Real-time exercise detection is a challenging problem. Hence, we present a real-time exercise detection system that addresses the challenge of accurately monitoring exercises without the need for an expensive, real-life coach. Our approach leverages machine learning techniques to detect errors in exercises and provide actionable feedback, making fitness training more accessible and affordable. First, we collect a dataset, capturing multiple video frames with joint landmarks represented as key points in each image. These images are labeled with exercise actions such as curls, presses, and squats. Then, we perform data cleaning and splitting. After that we have trained and evaluated two machine learning models: Long Short-Term Memory (LSTM) networks and Random Forest based on metrics, namely, accuracy, precision, recall, F1 score. From our experimental results, we have found that the LSTM model outperforms the Random Forest, demonstrating its superior effectiveness in real-time exercise detection. Our results highlight the potential of AI-driven systems to provide personalized fitness guidance, opening doors to innovative solutions in the health and fitness industry.