Deep learning-based physical exercise assessment of older adults using single-camera videos

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Geriatric Rehabilitation Technology

Background:

• Older adults in care homes often lack sufficient personalized supervision for physical activity, hindering functional independence.
• Technology-supported exercise platforms offer a potential solution, but require automated monitoring for safety and quality assurance.

Purpose of the Study:

• To design and evaluate a deep learning (DL) system for automated monitoring of rehabilitation exercises using standard video recordings.
• To assess the system's ability to recognize exercise types and estimate joint angle trajectories for objective performance metrics.

Main Methods:

• Developed a DL model using temporal convolutional neural networks to process 2D skeleton poses from video.
• Estimated 3D joint angles and recognized exercise types, calculating exercise performance metrics (EPMs) like duration and repetition count.
• Validated the system against ground-truth data from inertial sensors in seven care-home residents performing six common exercises.

Main Results:

• The DL model achieved high accuracy in exercise recognition (F1@50 of 0.92) and joint angle estimation (MPJAE of 7.7°).
• Estimated EPMs showed strong agreement with ground truth, with high correlations for duration (0.93) and repetition count (0.86).
• The system reliably estimated motion variability and range of motion across different exercises.

Conclusions:

• The developed DL algorithm effectively estimates key exercise outcomes from single video streams, enabling unsupervised assessment.
• This video-based monitoring pipeline can enhance exercise quality and safety in care homes without additional staff.
• Future research should focus on validating this approach in larger, diverse populations.