Validation of markerless joint-angle and range-of-motion measurement against 3D motion capture

Research question

Can joint angles and range of motion derived from markerless 2D human pose estimation agree closely enough with marker-based 3D motion capture to be useful for movement measurement?

Two pose-estimation models were compared, across several movements, to establish whether agreement depends on the model used, the movement performed, or both.

Population

Participants

22 healthy volunteers

Sex

16 female, 6 male

Mean age

36.9 ± 11.6 years

All participants were healthy adults.

Technology tested

MediaPipe and HRNet 2D human pose estimation. Joint angles were derived from body keypoint coordinates produced by each model.

• MediaPipe — a widely used real-time pose-estimation framework.
• HRNet — a high-resolution pose-estimation network.
• Joint angles calculated geometrically from the estimated keypoint positions.

Reference method

Qualisys marker-based 3D motion capture — the established laboratory reference standard for joint kinematics.

Protocol

1. 1Participants performed a set of movements in a laboratory instrumented with Qualisys marker-based 3D motion capture.
2. 2Movements included seated knee flexion/extension, prone knee flexion/extension, elbow flexion/extension and sit-to-stand.
3. 3Video of each movement was processed through both MediaPipe and HRNet pose estimation.
4. 4Joint angles were derived from the resulting keypoint coordinates and compared with the motion capture data.
5. 5Agreement was assessed for joint-angle time series using coefficient of variation, and for range of motion using intraclass correlation coefficients (ICC).

Results

All tested joint-angle time-series comparisons had coefficient-of-variation values below 10%.

Performance was both movement-specific and model-specific. Simple, clearly visible single-plane movements — seated knee flexion/extension and elbow flexion/extension — agreed most closely. Agreement was weaker for prone knee movement with HRNet, and weak and not statistically significant for left knee range of motion during sit-to-stand for both models.

Interpretation

Markerless pose estimation can reproduce joint-angle patterns and range of motion closely for some movements and less well for others. The choice of model matters, and so does how clearly the joint of interest is visible to the camera throughout the movement.

This supports selective, movement-specific use of markerless measurement rather than treating it as a general-purpose substitute for laboratory kinematics.

What this study validated

• That joint-angle time series derived from both models tracked the reference measurement with coefficient-of-variation values below 10% for every comparison tested.
• That range of motion for seated knee flexion/extension, prone knee flexion/extension, elbow flexion/extension and right knee sit-to-stand movement can show good-to-excellent agreement for at least one of the tested models.
• That agreement is movement-specific and model-specific, and should be established per movement.

Full citation

Hamilton RI, Glavcheva-Laleva Z, Haque Milon MI, Anil Y, Williams J, Bishop P, Holt C. Comparison of computational pose estimation models for joint angles with 3D motion capture. Journal of Bodywork and Movement Therapies. 2024;40:315–319.

DOI: 10.1016/j.jbmt.2024.04.033