Pose Estimation for Ergonomics: How Accurate Is Video-Based Assessment, and What Do Studies Show?

Why is accuracy the key question when you use pose estimation for ergonomics?

Pose estimation for ergonomics promises something attractive: point a phone at a task and get a REBA or RULA score in minutes. Whether that score is useful depends on how close the measured angles are to reality, and how well the final risk level matches what a trained assessor would conclude.

This post summarizes what peer-reviewed studies report on pose estimation for ergonomics, explains where errors come from, and shows how to test a tool on your own tasks. It applies to the whole category, including ViveLab's video ergonomic assessment, which produces an AI-assisted estimate processed on the device and is designed to be reviewed by a trained assessor.

What do we mean by accuracy?

Three different questions often get blurred together.

Term	Question it answers	Typical statistic
Reliability (intra-rater)	Does the same rater or tool give the same score twice?	Intraclass correlation (ICC)
Agreement (inter-rater)	Do two raters, or AI and a human, give the same score?	Kappa, weighted kappa, percent agreement
Validity (accuracy)	Is the measured angle or score correct against a reference?	Angle error in degrees versus motion capture

A tool can be perfectly reliable and still wrong in a consistent way. Agreement with a human is not the same as validity either, because human raters disagree with each other too.

What does the evidence say about reliability?

Here AI tools do very well. In a 2025 PLoS One study, three human raters (novice, intermediate and expert) scored REBA on 21 veterinary task videos three times each, at least a week apart, and a commercial deep-learning video tool scored the same videos. Intra-rater ICC for the final REBA score was:

Rater	Intra-rater ICC, final REBA score
AI tool	1.00
Expert	0.89
Intermediate	0.66
Novice	0.51

Inter-rater ICC among the three humans was only 0.36 (Balogh et al., 2025).

The takeaway is not that AI is better. It is that human REBA scoring varies a lot, especially with less experience, and software removes that source of noise.

How well does AI agree with expert raters?

Agreement depends heavily on the body segment. The same study compared the AI tool with each human rater using weighted kappa. Two conditions were tested: humans scoring the videos at timepoints they chose, and humans scoring still frames at the timepoints the AI selected.

Body segment	Expert vs AI, videos	Expert vs AI, AI-selected frames
Trunk	0.24	0.63
Legs	0.49	0.57
Upper arm	0.26	0.53
Neck	-0.27	0.18 (not significant)
Lower arm	0.06 (not significant)	0.23 (not significant)
Wrist	0.02 (not significant)	-0.09 (not significant)
Final REBA score	0.25	0.35

Source: Balogh et al., 2025.

Two patterns stand out. First, agreement jumps when humans and AI look at the same frames, which shows that a lot of disagreement comes from which moment is scored, not just how angles are measured. Second, large segments (trunk, legs, upper arm) agree far better than small or easily hidden ones (neck, lower arm, wrist).

The authors are clear about limits. They tested only three raters and REBA only, and they did not assess validity, so they cannot say whether the AI or the humans were more accurate. All views were side or oblique, which they suggest made wrist posture hard to judge.

How accurate are the joint angles themselves?

For validity, researchers compare camera-based systems against marker-based motion capture, the lab reference standard.

A 2025 systematic review in Sensors included 20 quantitative studies on industrial ergonomic risk analysis (Scataglini et al., 2025). It reported:

• Marker-based systems: highest precision (0.5 to 1.5 mm error) and reliability (ICC above 0.90), but limited by cost and lab constraints.
• Markerless camera-based systems: moderate-to-high accuracy (5 to 20 mm error; mean joint-angle error of 2.31 plus or minus 4.00 degrees) and good reliability (ICC above 0.80), with greater practicality in the field.
• Several studies reported strong validity for RULA and REBA prediction, with accuracy up to 89 percent and kappa of 0.71.

The review also warned that several studies used fewer than 10 participants, that reliability was underreported, and that some setups struggled with occlusion or self-occluded postures.

An earlier study built an OpenPose-based system for computing joint angles and RULA and REBA scores and validated it against a reference motion capture system. Recordings of 10 participants doing 12 tasks, with and without body occlusions and from frontal and non-frontal views, were analyzed. OpenPose showed good performance under all conditions, while a Kinect depth-sensor system performed significantly worse with occlusion or non-frontal tracking (Kim et al., 2021).

Note the scope: most of this evidence on pose estimation for ergonomics comes from lab-style tasks, small samples and specific camera setups. Results on a busy production floor with a handheld phone can differ.

Where do video-based errors come from?

Pose estimation for ergonomics fails in predictable ways, and understanding the error sources tells you how to reduce them.

2D projection

A single camera records a flat image. Movement toward or away from the lens shrinks on screen, so a trunk bent toward the camera can look less flexed than it is. Models such as MediaPipe Pose output estimated 3D world coordinates (Google AI Edge), but depth from one camera is still inferred, not measured.

Camera angle

Each view favors some angles. Side views capture flexion and extension; front views capture side bending and abduction. Twisting is hard from either. The Kim study's frontal versus non-frontal conditions, and the Balogh authors' comments on side-view wrist scoring, both point to view as a major factor.

Occlusion

When a joint is hidden by the body, a load, a machine or another person, the model has to infer where it is. The Sensors review flagged occlusion and self-occluded positions as a recurring problem.

Small joints

Wrist deviation, forearm rotation and grip are small movements. Full-body models track only a few points on each hand, and the hand is often behind the object being handled. This matches the near-zero wrist agreement in the REBA study.

Frame selection

REBA and RULA score a posture, usually the most demanding or most frequent one. If the AI and the assessor pick different moments, scores will differ even if both measure angles perfectly.

Inputs video cannot see

Load weight, force, coupling quality and activity level all change the final score. OSHA notes that weight alone does not determine lifting risk but that it is one of the factors in the NIOSH lifting equation (OSHA, 2013). No camera can read the weight of a closed box.

How can you improve accuracy when you record?

Error source	What to do
2D projection	Film perpendicular to the main movement, usually from the side for lifting
Camera angle	Add a second view for twisting or reaching tasks
Occlusion	Clear the line of sight, step back so the whole body is visible
Small joints	Add a close-up of the hands or score wrists manually
Frame selection	Review the frames the tool chose and override if they are not representative
Missing inputs	Enter load, force, coupling and activity yourself

How do you test a tool on your own tasks?

Published studies are a starting point. Your tasks, lighting and camera positions are what count.

1. Pick 10 to 20 tasks that cover your main job types, including at least a few with occlusion or twisting.
2. Record each one following the recording tips above.
3. Have a trained assessor score them manually, without seeing the AI result. Our REBA assessment guide and RULA vs REBA comparison help you pick the method.
4. Run the AI tool on the same videos.
5. Compare by body segment and by final risk level, not just the final number.
6. Look at disagreements. If they cluster on wrists or twisting tasks, adjust your workflow so those are always checked manually.

What you are looking for is whether the tool reliably separates jobs that need action from jobs that do not.

So how should you use video-based assessment?

Use it for what the evidence supports:

• Screening many jobs quickly to find likely high-risk ones.
• Consistent re-measurement before and after a fix, since the same tool on the same setup will not drift the way human raters do.
• Documentation, with frames and angles attached.

Keep a trained assessor on the decisions: checking angles, adding missing inputs, confirming high scores and choosing controls. That is how ViveLab is designed to be used. It gives an AI-assisted estimate processed on the device, and a person signs off. For a broader look at the tool landscape, read our overview of video-based ergonomic assessment software.

Next steps

The best test of pose estimation for ergonomics is your own work. Run one of your own tasks through AI posture analysis, then score the same task by hand and compare segment by segment. Our guide to choosing a posture analysis app has a recording checklist to get the cleanest input.