Interactive Exercise Guidance with Real-time Pose Correction Using Computer Vision
Real-time Posture Analysis & Feedback Systems
60
Training Videos
18
Validated Videos
5
Test Participants
Monocular
Input Type
Overview
This system aims to provide real-time feedback for physical exercises using standard monocular camera setups. It combines geometric rule-based analysis with a machine-learning posture predictor (PosePredictor) to identify faulty postures and provide corrective hints to users. The current research focuses on the limitations of small-scale datasets and monocular depth ambiguities in real-world fitness applications.
Methodology
The system was designed around a fixed set of exercises including overhead press, bicep/lateral curls, and push-ups. Form feedback logic is bespoke to each exercise, based on expected joint angle cycles. Due to the monocular setup, the system relies on temporal (frame-to-frame) context to resolve depth ambiguities and uses sequence smoothing for rotation correction.
Impact & Applications
The research highlights the transition from purely rule-based systems to hybrid models. By using ML outputs as auxiliary signals (triggering "uncertainty" hints), the system maintains reliability even with limited data. This foundational work informs future developments in synthetic data augmentation and automated exercise counting across varied body types and environments.
Known Limitations & Generalizability
The machine-learning components were trained on a limited dataset of 60 exercise videos, with only 18 fitting high-quality criteria, making the model prone to overfitting. Evaluation involved a small test cohort of five participants, assuming body types similar to the test group. The lack of multi-view or depth data means the system relies solely on 2D monocular input, which struggles with complex depth ambiguities without temporal smoothing.
Keywords