Advancing functional reach assessments: a comparison of YOLOv8 human pose estimation and 3D motion capture in a young healthy cohort

Accurate assessment of dynamic sitting balance is essential in rehabilitation. Traditional 3D motion capture systems provide precise measurements but are costly and impractical for daily clinical use. Human Pose Estimation (HPE) offers a low-cost and less time-consuming alternative to clinical assessments for Spinal Cord Injuries compared to traditional measurement techniques. Sixteen healthy adults performed forward and lateral Modified Functional Reach Tests across two sessions. Movements were recorded using an 8-camera 3D motion capture system and two 2D cameras. YOLOv8n (lightweight version) and YOLOv8x (heavy version) HPE models were applied to estimate reach distances. Reliability was assessed using intraclass correlation coefficients (ICC), standard error of measurement (SEM), and coefficient of variation (COV). Validity was evaluated against 3D motion capture using repeated measures correlation and Bland–Altman analysis. YOLOv8n showed excellent reliability and strong concurrent validity for forward reaching (ICC ≥ 0.98; rm-R2 = 0.98), while YOLOv8x performed better in the lateral reaching task (ICC ≤ 0.35; rm-R2 = 0.73). Findings suggest both models underestimated lateral reach distances compared to 3D motion capture but a practical and reliable solution for forward reach assessment and shows potential for clinical application in resource-limited settings. However, reduced reliability and spatial agreement in lateral reaching highlights the need for further refinement and validation in clinical populations. This study provides a foundation for integrating deep learning-based pose estimation into rehabilitation practice, enabling accessible, markerless motion analysis for dynamic balance evaluation.