Evaluating AI-assisted motion tracking for knee rehabilitation: a comparative analysis of You Only Look Once version 8 (YOLOv8) and MediaPipe

Conventional methods for measuring joint angles and body movement, such as goniometers and inertial measurement units (IMUs), can be impractical for home-based use due to the need for physical setup, calibration, or supervision. This has led to a growing interest in AI-based, markerless pose estimation techniques as a more convenient alternative. This study compares two real-time AI-powered computer vision models MediaPipe and You Only Look Once (YOLOv8), specifically the YOLOv8n and YOLOv8s variants for their reliability, error percentage, and inference time in tracking knee joint angles during three rehabilitation exercises: Seated Knee Extension, Half Squat, and Sit-to-Stand. During the rehabilitation exercises, the knee was connected to a goniometer and performed by one participant while another person supervised to compare its readings with computer-based measurements. Each exercise set was repeated four times under supervision. The objective of this study is to develop a non-invasive and low-cost motion tracking framework suitable for remote rehabilitation, aiming to improve patient monitoring and exercise compliance. MediaPipe outperformed both YOLOv8 models, demonstrating the lowest error at full extension (180 ) at 0.7779% and the fastest inference time (approximately 35ms). YOLOv8n performed well at 90 flexion (error: 0.05%) but showed a higher error at 180 (4.94%). YOLOv8s showed the highest error at 180 (8.87%) due to landmark misalignment. Inference times were also significantly higher for YOLOv8n (183.79 ms) and YOLOv8s (291.34 ms), compared to MediaPipe (36.45ms), with increased noise and jitter during dynamic motion. Overall, MediaPipe's lightweight and Central Processing Unit (CPU)-optimized design makes it the most suitable for real-time, non-invasive, and low-cost motion tracking in remote rehabilitation, while the YOLOv8 models offer better spatial detail at the expense of speed and stability.

Text 123549.pdf - Published Version Restricted to Repository staff only Download (764kB) | Request a copy Official URL or Download Paper: https://ieeexplore.ieee.org/document/11258043/

Actions (login required)

View Item