💻 computer science

Monocular Markerless Motion Capture Enables Quantitative Assessment of Upper Extremity Reachable Workspace

This study validates the feasibility of using a single frontal camera with AI-driven markerless motion capture to quantitatively assess the upper extremity reachable workspace, demonstrating strong agreement with traditional marker-based systems and offering a clinically accessible alternative for biomechanical analysis.

Original authors: Seth Donahue, J. D. Peiffer, R. Tyler Richardson, Yishan Zhong, Shaun Q. Y. Tan, Benoit Marteau, Stephanie R. Russo, May D. Wang, R. James Cotton, Ross Chafetz

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seth Donahue, J. D. Peiffer, R. Tyler Richardson, Yishan Zhong, Shaun Q. Y. Tan, Benoit Marteau, Stephanie R. Russo, May D. Wang, R. James Cotton, Ross Chafetz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to measure how far a person can reach with their arm. In the past, to do this accurately, you needed a high-tech "lab in a room." You'd have to tape dozens of tiny, shiny stickers (markers) all over the person's arm and back, set up a dozen expensive cameras, and hire a team of engineers to make sure everything was perfectly calibrated. It was like trying to film a movie with a Hollywood crew just to see if someone could touch a light switch.

This paper introduces a much simpler, cheaper, and faster way to do the same thing: using just one regular camera and some smart computer software (Artificial Intelligence).

Here is the breakdown of what they did, how it worked, and what they found, using some everyday analogies.

The Goal: Mapping the "Reach Zone"

Think of your arm's movement as a giant, invisible bubble around your body. Some parts of that bubble are easy to reach (like touching your nose), while others are hard (like reaching behind your back or way over to the other side). Doctors call this the Reachable Workspace.

To help patients recovering from strokes or injuries, doctors need to know exactly how big that bubble is and if it's shrinking or growing. The problem is, the old "sticker-and-camera" method is too expensive and complicated for a regular doctor's office.

The Experiment: The "One-Camera" Test

The researchers wanted to see if a single camera could do the job of the expensive 12-camera system.

The Setup: They asked 9 healthy adults to play a video game in a Virtual Reality (VR) headset. The game showed them virtual targets floating in the air around them. The goal was to reach out and "touch" as many targets as possible with their right hand.
The Comparison: While the people played, the researchers recorded them in two ways:
- The Gold Standard: The old-school way with 12 cameras and sticky markers (the "Hollywood crew").
- The New Way: A single camera using AI to guess where the joints are without any stickers.
The Twist: They tested the single camera in two positions:
- Frontal: Directly in front of the person (like a webcam on a laptop).
- Offset: Angled to the side (like someone standing at a 45-degree angle).

The Results: Front vs. Side

The researchers compared the "One-Camera" results against the "Gold Standard" to see how close they were.

1. The Frontal Camera (The "Face-to-Face" View)

The Analogy: Imagine looking at someone straight on. You can clearly see if they are reaching forward or to the side.
The Result: This was a huge success. The single camera sitting in front of the person gave results almost identical to the expensive 12-camera system.
Why it works: When you look at someone from the front, their arm moves across your view (left to right). The AI can easily track that movement. It was accurate enough to be used in a real doctor's office.

2. The Offset Camera (The "Side-View" View)

The Analogy: Imagine trying to judge how far someone is reaching away from you while you stand to their side. It's hard to tell if their hand is right in front of them or way back behind them.
The Result: This camera underestimated how far people could reach. It thought the patients could reach less than they actually could.
Why it failed: When the arm moves "into" the camera (towards or away from the lens), it's hard for a single 2D camera to guess the depth. It's like trying to guess how far away a car is just by looking at its headlights; it's easy to get the distance wrong. Also, if the person's body blocked the view of their hand, the AI got confused.

The "Depth" Problem

The main issue with using just one camera is Depth Perception.

Binocular Vision (Two Eyes/Cameras): Humans have two eyes, so we can judge depth perfectly.
Monocular Vision (One Eye/Camera): A single camera is like closing one eye. It can see left/right and up/down easily, but it struggles to guess "forward/backward."
The Fix: The researchers found that if you position the camera so the person is reaching across the screen (not into the screen), the AI can guess the depth well enough to be useful.

Why This Matters

This study is a game-changer for two reasons:

Accessibility: You don't need a million-dollar lab anymore. A doctor could potentially use a tablet or a simple webcam to assess a patient's arm mobility in a regular clinic, or even let a patient do it at home.
Speed: No more taping stickers on patients. No more waiting for a team to set up 12 cameras. Just hit record, play the game, and get the data.

The Catch (The "But...")

While the frontal camera is great for seeing how far people can reach forward (which is what most patients need to do for daily tasks like eating or brushing teeth), it still struggles a bit with reaching behind the back or deep into the corners of the room.

Think of it like a new smartphone camera: It takes amazing photos of your face, but if you try to take a photo of something far away in the dark, it might get a little blurry. The technology is good enough to be useful now, but it needs a little more polishing to be perfect for every single angle.

The Bottom Line

This paper proves that one camera is enough to measure how well a patient can use their arm, as long as you position the camera correctly (in front of them). It turns a complex, expensive science experiment into something as simple as taking a video, making high-quality medical analysis available to everyone, not just those in fancy research labs.

1. Problem Statement

Quantitative assessment of Upper Extremity (UE) mobility is critical for clinical diagnosis and rehabilitation tracking. The Upper Extremity Reachable Workspace (UERW) task, which measures the volume of space a patient can reach, is a validated clinical metric. However, the current "gold standard" for measuring UERW relies on marker-based motion capture (MoCap) systems. These systems face significant barriers to clinical adoption, including:

High equipment costs and the need for specialized laboratory spaces.
Complex setup requiring multiple synchronized cameras and calibration.
Labor-intensive preparation (applying retro-reflective markers) and data processing.
Lack of scalability for point-of-care or home-based rehabilitation.

While Markerless Motion Capture (MMC) using Artificial Intelligence (AI) offers a solution, existing high-accuracy MMC systems often require multi-camera setups (e.g., Theia3D, OpenCap), which still suffer from logistical complexity. Single-camera (monocular) MMC is more scalable but faces challenges with depth estimation errors and anatomical occlusion, particularly for dynamic, multi-planar movements like the UERW task. There is a lack of validation for monocular MMC in specific clinical UERW assessments compared to marker-based gold standards.

2. Methodology

The study aimed to validate a monocular MMC system against a marker-based reference for the UERW task.

Participants: Nine healthy adults (5 males, 4 females) with no upper extremity impairments.
Experimental Setup:
- Task: Participants wore a Meta Quest 3 VR headset and reached for virtual targets distributed across a sphere around their torso. Targets disappeared upon reaching within 5cm.
- Ground Truth: Simultaneous data collection using a 12-camera marker-based Vicon system (60 Hz) with retro-reflective markers on the wrist, scapula, and torso.
- Test Systems: Two single-camera configurations using FLIR Blackfly S cameras:
  1. Frontal View: Positioned anterior and superior to the participant.
  2. Offset View: Positioned anterior, superior, and lateral (~45° to the right of the frontal view).
Data Processing Pipeline:
- Keypoint Detection: Used PosePipe with MeTRAbs-ACAE to extract 2D/3D anatomical keypoints (MoVi set of 87 keypoints).
- Biomechanical Reconstruction: A whole-body biomechanical model was fitted to the keypoints using MuJoCo (a physics engine). The system utilized a learned implicit function (multilayer perceptron) to optimize joint angles ( $\theta$ ) and body scaling parameters ( $\beta$ ) to minimize 2D and 3D loss functions.
- Coordinate Alignment: MMC coordinate systems were aligned with the marker-based torso coordinate system to ensure comparable metrics.
Metrics:
- Primary Outcome: Percentage of workspace reached (% Reachspace) across six specific octants (e.g., Superior Anterior Ipsilateral, Inferior Posterior Ipsilateral).
- Secondary Outcomes: Peak reach distance differences, agreement rates (octant identification), and directional error rates (Anterior-Posterior, Superior-Inferior, Medial-Lateral).
Statistical Analysis: Two-way repeated-measures ANOVA, Bland-Altman analysis, and pairwise comparisons with Bonferroni correction.

3. Key Contributions

First Validation: This is the first study to validate a monocular, markerless system specifically for the dynamic, multi-planar UERW clinical task against a marker-based gold standard.
Configuration Comparison: It systematically compares frontal vs. offset camera views, identifying that camera orientation significantly impacts depth estimation and occlusion handling in clinical tasks.
Clinical Feasibility: Demonstrates that a single-camera setup can achieve clinical-grade accuracy for anterior workspace assessment, potentially democratizing access to quantitative motion analysis.
Error Characterization: Provides a detailed breakdown of where and why monocular systems fail (e.g., posterior depth ambiguity and contralateral occlusion), offering a roadmap for future algorithmic improvements.

4. Results

Peak Reach Distance: Differences between systems were minimal ( $0.04 \pm 0.03$ m), indicating the reconstruction pipeline accurately estimated limb length and reach limits.
Workspace Agreement (Frontal vs. Marker):
- The frontal camera showed strong agreement with the marker-based system.
- Mean bias was minimal ( $0.61 \pm 0.12\%$ ).
- No statistically significant differences were found in the percentage of workspace reached across any of the six octants ( $p > 0.303$ ).
- High agreement rates (>90%) were observed in anterior octants.
Workspace Agreement (Offset vs. Marker):
- The offset camera significantly underestimated the workspace reached (mean bias: $-5.66 \pm 0.45\%$ ).
- Significant differences were found in the Superior Anterior Contralateral and Inferior Anterior Contralateral octants.
- Agreement dropped to <40% in the Superior Anterior Contralateral octant due to torso/arm occlusion.
Error Sources:
- Frontal View: Errors were primarily confined to posterior octants due to depth ambiguity (foreshortening along the optical axis).
- Offset View: Errors occurred in both contralateral (due to occlusion by the torso) and posterior octants.
- Directional errors were highest in the Anterior-Posterior (AP) and Medial-Lateral (ML) axes when movement vectors aligned with the camera's optical axis or were obstructed.

5. Significance

Clinical Adoption: The study proves that a frontal monocular camera is a viable, low-cost alternative to complex multi-camera labs for assessing UE mobility, particularly for anterior workspace evaluation (critical for stroke and brachial plexus palsy rehabilitation).
Scalability: By eliminating the need for multiple cameras, calibration, and marker application, this approach enables quantitative assessment in point-of-care settings and home-based rehabilitation.
Limitations & Future Work:
- Current accuracy is limited in posterior and contralateral spaces due to depth estimation and occlusion.
- The study was conducted on healthy adults; future work must validate performance on clinical populations with atypical kinematics (e.g., tremors, compensatory movements).
- Future iterations should focus on real-time feedback integration and mobile platform deployment to transition from post-hoc analysis to immediate clinical utility.

In conclusion, this research establishes that AI-driven monocular MMC can effectively replace marker-based systems for specific, high-value clinical assessments, significantly lowering the barrier to entry for objective biomechanical analysis.