FURI | Spring 2025

Machine Learning for Exercise Form Correction using Pose Detection and Camera Tracking

The research collects labeled exercise data (correct vs. incorrect form), uses HOG (Histogram of Oriented Gradients) to extract features, and leverages Google’s MediaPipe framework to identify key pose landmarks. This LSTM-based model captures sequential patterns to detect specific deviations across each repetition; by providing immediate feedback, the system aims to reduce gym-related injuries, streamline fitness progress, and support future expansion into 3D kinematics, following the proposed steps for data collection, model refinement, and real-world validation.