Author(s): SHIU, K., ONG, A., GAO, X., Institution: REPUBLIC POLYTECHNIC, Country: SINGAPORE, Abstract-ID: 2250

INTRODUCTION:
Conventional three‑dimensional (3D) motion capture (MoCap) systems are the gold standard for gait analysis but are costly, time‑intensive, and confined to laboratory settings. Advances in machine learning (ML) and computer vision have enabled markerless gait analysis using a single camera; however, validation against gold‑standard systems remains limited. This study evaluated the accuracy of a single‑camera, ML‑based approach (MediaPipe Pose) for estimating 3D joint kinematics during gait, compared with a laboratory‑based MoCap system.
METHODS:
Eleven healthy young adults each completed 30 walking trials at slow, normal, and fast speeds. Kinematic data were recorded simultaneously using a multi‑camera MoCap system and a single smartphone camera. MoCap marker trajectories were processed and joint forces and moments were computed using inverse dynamics. For the single-camera ML-based approach, MediaPipe Pose extracted 3D spatial joint data directly from synchronized video recordings. Joint linear kinematics were derived along the X, Y, and Z axes. Agreement between systems was assessed using cross‑correlation analysis, and differences in kinematic magnitudes across measurement methods and walking speeds were examined using repeated‑measures ANOVA (p < 0.05).
RESULTS:
Strong agreement was observed primarily for X‑axis kinematics. Cross‑correlation analysis demonstrated strong waveform similarity for several joints, including the knee (r = 0.738–0.833) and ankle (r = 0.827–0.844), with minimal temporal lag. Several X‑axis variables showed no significant differences between systems (e.g., left elbow p = 0.973; left wrist p = 0.231). In contrast, Y‑ and Z‑axis kinematics showed reduced agreement, with weak correlations (e.g., knee Z: r = 0.165–0.219) and significant differences for the hip, knee, and ankle (p ≤ 0.001). Waveform inspection revealed smooth, symmetric X‑axis trajectories, whereas Y‑ and Z‑axis signals exhibited reduced amplitude, increased noise, and timing offsets.
CONCLUSION:
The single‑camera ML‑based approach demonstrates high validity for sagittal‑plane gait kinematics associated with forward progression and joint flexion–extension, with several X‑axis variables producing statistically similar results between systems.
In contrast, reduced agreement in Y‑ and Z‑axis measures likely reflects known limitations of single‑camera ML approach in capturing depth and out‑of‑plane movements such as abduction/adduction and rotational components of gait. While laboratory‑based MoCap remains essential for comprehensive 3D biomechanical analysis, single‑camera ML approach shows strong potential for accessible sagittal‑plane gait assessment, screening, and field‑based research.