RoMed: Webcam-Based ROM Assessment

• RoMed is a markerless range-of-motion evaluation tool that uses a single RGB webcam stream to estimate joint movement with high repeatability.
• It leverages MediaPipe BlazePose to detect real-time body landmarks, providing a practical alternative to manual goniometry and costly motion capture.
• Experimental results demonstrate substantial test-retest reliability and effective telehealth deployment despite challenges in capturing maximal flexion angles.

RoMed is a webcam-based range-of-motion evaluation method proposed as a machine learning–based, markerless system for estimating joint range of motion from a single RGB webcam stream. In the formulation reported in "A webcam-based machine learning approach for three-dimensional range of motion evaluation" (Wang et al., 2023), RoMed is built on MediaPipe BlazePose and is positioned as a practical alternative to goniometry and marker-based optical motion capture for remote physical therapy and rehabilitation. Its central function is to detect body landmarks in real time, reconstruct approximate 3D joint positions, and compute movement excursion relative to a baseline posture.

1. Definition, scope, and clinical rationale

RoMed addresses a specific measurement problem in rehabilitation: quantitative assessment of joint range of motion (ROM), a core outcome in physical therapy. The paper situates the method against two established baselines. First, goniometry is widely used but requires training and is vulnerable to human error. Second, optical motion capture is accurate but expensive and impractical for routine care. RoMed is introduced as a computer-vision alternative that can be accessed remotely through a webcam-equipped device (Wang et al., 2023).

In this context, RoMed is not a generic reduced-order modeling term; it refers specifically to a markerless ROM assessment tool for physical therapy. The method is designed for home and telehealth use, where patients may lack in-person access to trained clinicians, goniometers, or laboratory-grade motion capture systems. This positioning is clinically significant because it links algorithmic pose estimation to a concrete access problem in rehabilitation delivery.

The study is explicitly framed as a proof-of-concept reliability study. Its objective is not to replace all clinical ROM workflows unconditionally, but to evaluate whether webcam-derived measurements are sufficiently reliable to support clinical practice and tele-implementation of physical therapy and rehabilitation (Wang et al., 2023).

2. Experimental setting and movement coverage

The evaluation dataset comprised 25 healthy adults, including 12 males and 13 females, recruited from the University of Wyoming community. Data collection occurred from March to November 2022. All participants gave written informed consent, and the study was approved by the University of Wyoming IRB (Wang et al., 2023).

The movement set was deliberately broad, spanning the spine, neck, upper extremity, and lower extremity. Distal joints such as the wrist and ankle were excluded because BlazePose could not reliably track hands and feet in the reported setup.

Region	Movements evaluated
Spine	Back flexion and extension; back lateral flexion; trunk rotation
Neck	Flexion and extension; lateral bending; rotation
Upper extremity	Shoulder adduction and abduction; shoulder flexion and extension; elbow flexion
Lower extremity	Hip flexion and extension; hip flexion (knee flexed); hip adduction and abduction

This movement selection matters methodologically because it tests RoMed across different kinematic regimes: axial motion, large-amplitude limb motion, and movements prone to self-occlusion. A plausible implication is that the reported performance is shaped not only by landmark localization accuracy but also by movement geometry and visibility constraints.

3. Computational pipeline and mathematical formulation

The RoMed pipeline begins with video capture from a Logitech C922 Pro HD Stream webcam positioned about 4 m from participants at approximately chest height. Participants were oriented to reduce occlusion according to the movement being measured (Wang et al., 2023).

Pose estimation is performed with Google MediaPipe’s BlazePose. In the reported configuration, BlazePose detects a person in the frame, tracks that person across frames, outputs 33 body landmarks in 3D, and in principle runs at over 30 FPS. Because of synchronous streaming constraints, the effective rate in the study was about 15 FPS. Each landmark also carries a visibility index on the interval from 0 to 1; RoMed used a threshold of 0.5, and landmarks below that threshold were removed (Wang et al., 2023).

ROM angles are computed from selected landmarks or derived segment points. The paper gives examples of how these segments were defined: back flexion/extension used hip and shoulder landmarks for BlazePose, while neck flexion/extension used the nose and midpoint of the shoulders. For OptiTrack comparison, some joint definitions required averaging markers to form equivalent segment points.

The segment orientation is normalized as

$v = \frac{P_1 - P_2}{\lVert P_1 - P_2 \rVert}.$

The movement angle at time $t$ is then computed as

$a_t = \arccos(v_t \cdot v_0),$

where $v_0$ is the segment orientation at the starting position and $v_t$ is the segment orientation in the current frame. The final ROM estimate is defined as the maximum angle during the movement (Wang et al., 2023).

A notable feature of the implementation is that the raw maximum is not accepted uncritically. To identify the correct maximum in the presence of tracking noise, the paper applies additive seasonality decomposition. The trend is taken as the movement-angle trajectory, the residual is used to detect anomalies, and outliers are defined as residuals 3 standard deviations from the mean. This choice indicates that RoMed is not merely a landmark tracker; it is a full measurement pipeline with explicit temporal post-processing for peak detection (Wang et al., 2023).

4. Reference system and reliability methodology

RoMed was evaluated against a marker-based infrared optical motion capture system treated as the ground-truth comparator. The reference platform was OptiTrack with 10 Primex 22 cameras, capturing motion at 120 FPS and using a 39-marker full-body biomechanical model (Wang et al., 2023).

The study examined three forms of reliability: test-retest reliability within BlazePose, test-retest reliability within OptiTrack, and between-method reliability between BlazePose and OptiTrack. The reported metrics were intraclass correlation coefficient (ICC), standard error of measurement (SEM), minimal detectable change (MDC), and linear regression $r^2$ values (Wang et al., 2023).

For ICC interpretation, the paper uses the following scale:

• $0$–$0.2$: slight
• $0.2$–$0.4$: fair
• $t$0–$t$1: moderate
• $t$2–$t$3: substantial
• $t$4–$t$5: almost perfect

The ICC models were a two-way mixed-effect model for multiple measurements for test-retest reliability and a two-way mixed-effect ICC for between-method reliability. The paper gives the following formulas:

$t$6

and

$t$7

Within the study, SEM estimates measurement error in degrees, and MDC represents the smallest change likely to exceed random error at the 95% confidence level (Wang et al., 2023).

5. Reliability results and performance profile

The principal empirical result is that RoMed showed high test-retest reliability across most examined joints. BlazePose-derived measurements were reported as substantial to almost perfect for most movements, with many ICCs in the $t$8 range, SEM often below about $t$9, and MDC often around $a_t = \arccos(v_t \cdot v_0),$0 or less for many tasks (Wang et al., 2023).

Representative BlazePose test-retest ICCs included back flexion at $a_t = \arccos(v_t \cdot v_0),$1, neck flexion at $a_t = \arccos(v_t \cdot v_0),$2, shoulder abduction at $a_t = \arccos(v_t \cdot v_0),$3, and hip abduction at $a_t = \arccos(v_t \cdot v_0),$4. Lower but still substantial values were reported for elbow flexion at $a_t = \arccos(v_t \cdot v_0),$5 and hip flexion (knee flexed) at $a_t = \arccos(v_t \cdot v_0),$6 (Wang et al., 2023).

The OptiTrack comparator was also generally reliable, but the study reports some unexpectedly lower values attributable to marker instability or occlusion in the experimental setup. Back extension had a very low ICC of $a_t = \arccos(v_t \cdot v_0),$7, while back flexion was $a_t = \arccos(v_t \cdot v_0),$8; many other movements remained in the substantial-to-almost-perfect range (Wang et al., 2023).

Between-method agreement was more variable. High agreement examples included back lateral flexion ($a_t = \arccos(v_t \cdot v_0),$9), trunk rotation ($v_0$0), neck flexion ($v_0$1), shoulder adduction ($v_0$2), shoulder extension ($v_0$3), hip adduction ($v_0$4), and hip abduction ($v_0$5). Lower agreement was observed for shoulder flexion ($v_0$6), elbow flexion ($v_0$7), and hip flexion with the knee flexed ($v_0$8) (Wang et al., 2023).

The regression analysis followed the same pattern: movements with strong between-method agreement had higher $v_0$9, while poorly aligned movements had low $v_t$0, most notably shoulder flexion with $v_t$1. Taken together, these findings distinguish two different properties of the system. RoMed is highly repeatable within itself for many tasks, yet agreement with marker-based motion capture depends strongly on the specific movement being measured. This suggests that repeatability alone is insufficient to characterize performance in markerless ROM assessment.

6. Failure modes, limitations, and interpretive issues

The central limitation identified in the paper is reduced sensitivity to joint location at the apex of movement, especially for shoulder flexion, elbow flexion, and hip flexion with the knee flexed (Wang et al., 2023). At maximal flexion, several factors can degrade landmark estimation: body segments overlap, contrast may be reduced by the all-black motion-capture suit, estimated joint positions can deviate from the actual positions, and the resulting trajectory may appear flat near the peak rather than parabolic.

This limitation is important because it clarifies an otherwise potentially misleading pattern in the results. A movement can exhibit high test-retest reliability within the webcam system while still showing weak agreement with OptiTrack. In RoMed, the problem is not necessarily random instability; it can be systematic loss of sensitivity near the movement apex. That distinction matters for both validation and deployment, since it implies that certain kinematic classes are intrinsically more difficult for a single-camera pose-estimation pipeline.

Another practical limitation is coverage. Distal joints were excluded because BlazePose could not reliably track hands and feet in the study configuration (Wang et al., 2023). The evaluated method is therefore best understood as a markerless ROM tool for selected spinal, cervical, proximal upper-extremity, and proximal lower-extremity movements, rather than a universal whole-body kinematic measurement system.

7. Clinical and technological significance

The paper concludes that the proposed webcam-based method exhibited high test-retest and inter-rater reliability, making it a versatile alternative for existing ROM evaluation methods in clinical practice and the tele-implementation of physical therapy and rehabilitation (Wang et al., 2023). The most direct implication is for tele-rehabilitation: RoMed could allow remote ROM assessment using only a webcam-equipped device, thereby reducing dependence on in-person measurement by trained clinicians.

The study also notes that high intra-rater reliability implies usefulness for longitudinal monitoring. In practical terms, changes greater than about $v_t$2 are likely to exceed measurement error for many movements (Wang et al., 2023). This does not imply equivalence to laboratory motion capture across all tasks, but it does indicate that the system may be sufficient for detecting clinically meaningful changes over time in many use cases.

The authors further note that the system could eventually be deployed as a browser-based or mobile application. This suggests a pathway from proof-of-concept reliability testing toward scalable telehealth instrumentation. A plausible implication is that RoMed occupies an intermediate space between manual goniometry and full motion laboratories: less resource-intensive than the latter, less operator-dependent than the former, and especially relevant where access barriers are the dominant constraint.

In summary, RoMed denotes a single-camera, BlazePose-driven, markerless ROM evaluation method whose main strength lies in high repeatability and practical deployability for remote rehabilitation. Its empirical profile is not uniform across motions: performance is strong for many neck, spine, shoulder, and hip tasks, but weaker for some flexion movements at maximal range because of apex sensitivity and landmark-tracking limitations (Wang et al., 2023).