Comparative evaluation of single and multiple camera-based pose estimation methods for gait spatiotemporal parameter estimation
Figures
Abstract
Single-camera pose estimation may provide a more accessible approach to measure a person’s walking patterns (i.e., gait analysis) than multi-camera markerless motion capture (MMC), but its validity and test-retest reliability remain unclear. This study evaluated spatiotemporal gait parameters from a single iPad camera against an eight-camera MMC reference system across four visits, two weeks apart. Thirty-six healthy adults (18.7 ± 1.7, ranging 18‒27 years) completed usual-gait, bag-carrying, and phone-use walking conditions. Validity analyses included 458 paired trials from Visit 1 and used repeated-measures and mixed-effects methods. Reliability analyses included 33 participants with complete four-visit data and evaluated absolute agreement of single visit-level measurements. Using repeated-measures correlation, the strongest within-participant association between systems was gait speed (rrm = 0.588), followed by stance duration (rrm = 0.528), cycle time (rrm = 0.515), and cadence (rrm = 0.446; all p < 0.001). Step length, calculated as half of stride length, and stride length showed weaker associations (rrm = 0.327 and 0.328; both p < 0.001), whereas swing duration was not significantly associated (p = 0.599). Independent step length and left−right step-length asymmetry were not evaluated. Mean system differences were small, but significant proportional bias was present for all seven gait variables, indicating that system differences changed with measurement magnitude. Walking condition influenced bias for cadence, gait speed, cycle time, step length, and stride length. Test-retest reliability varied by parameter and condition, with absolute-agreement intraclass correlation coefficients ranging from 0.499 to 0.862. Gait speed demonstrated good reliability across all three conditions, step and stride length demonstrated their highest reliability during the phone condition, and swing duration consistently demonstrated the lowest reliability. These findings support the feasibility of single-camera MMC for selected spatiotemporal gait outcomes, particularly gait speed, when accessibility and reduced equipment requirements are priorities. However, the approaches should not be considered interchangeable. Multi-camera MMC may remain better suited when three-dimensional reconstruction or more comprehensive spatial and temporal characterization is required.
Citation: Weber B, Schoenwether B, Booth T, McFarland M, Britton J, Eltoukhy M (2026) Comparative evaluation of single and multiple camera-based pose estimation methods for gait spatiotemporal parameter estimation. PLoS One 21(11): e0363369. https://doi.org/10.1371/journal.pone.0363369
Editor: Jyotindra Narayan, Indian Institute of Technology Patna, INDIA
Received: September 9, 2026; Accepted: November 7, 2026; Published: November 20, 2026
Copyright: © 2026 Weber et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files.
Funding: The author(s) received no specific funding for this work.
Competing interests: NO authors have competing interests.
Introduction
Gait analysis is widely used to quantify the spatial and temporal characteristics of walking and evaluate functional movement across clinical and research settings [1–5]. Gait analysis provides objective information that supplements observation and may contribute to diagnosis, treatment planning, and monitoring of functional change over time [1–5]. Commonly reported gait outcomes include spatiotemporal parameters, such as‒ gait speed, cadence, step length, stride length, gait cycle time, and stance and swing durations [3,4]. These measures characterize the timing and distance components of walking and have demonstrated sensitivity to functional impairments associated with conditions such as knee osteoarthritis and multiple sclerosis [6,7]. Gait characteristics may also vary with psychological and cognitive state, further illustrating the sensitivity of spatiotemporal measures to changes in human function [8,9].
The objective assessment of gait relies on a variety of motion-analysis technologies, each with distinct strengths and limitations that have driven the development of newer approaches. These technologies include instrumented walkways, wearable inertial sensors, and optical motion-capture systems [3,4,10]. Multi-camera optical marker-based motion capture is widely regarded as a reference method for biomechanical movement analysis because it permits three-dimensional reconstruction of body-segment motion [3–5]. However, marker-based systems require specialized laboratory infrastructure, participant preparation, and accurate placement of reflective markers on anatomical landmarks. Measurement may also be affected by marker-placement variability and movement of markers relative to the underlying skeletal structures [1,3–5]. These limitations have contributed to increasing interest in markerless motion capture (MMC) approaches that estimate human movement directly from video.
Advances in computer vision and artificial intelligence have enabled MMC systems to identify anatomical landmarks from video and reconstruct human movement without physical markers [10,11]. Multi-camera MMC systems use synchronized views to reconstruct three-dimensional body trajectories and have demonstrated promising validity for spatiotemporal gait analysis in healthy adults and clinical populations [10,12–15]. Previous studies using artificial intelligence driven multi-camera MMC have reported favorable agreement with established reference systems for variables including gait speed, stride length, step length, and several temporal parameters [12,14,15]. Systematic reviews similarly indicate that spatiotemporal gait measures are among the more accurately estimated variables obtained using contemporary markerless approaches [10,11]. Reliability is also necessary if markerless systems are to be used for repeated gait assessment. Previous studies have demonstrated acceptable to excellent test-retest reliability for multiple spatiotemporal parameters obtained using multi-camera MMC [16,17]. For example, Schoenwether et al. [17] demonstrated high between-session reliability for gait measures derived from an eight-camera artificial intelligence-driven MMC system in healthy adults, while Riazati et al. [16] reported favorable absolute reliability for gait parameters collected outside conventional motion-analysis laboratories using an MMC system. Collectively, these findings support the use of multi-camera MMC for gait analysis; however, such systems still require multiple synchronized cameras, dedicated capture volumes, and relatively fixed acquisition configurations.
Single-camera MMC offers a potentially more accessible alternative to synchronized multi-camera systems. Conventional single-camera (RGB) video, a single camera provides two-dimensional image-plane information from which anatomical landmarks can be estimated using computer-vision-based pose-estimation methods. The resulting two-dimensional landmark trajectories can then be used to identify gait events and derive planar spatiotemporal measures without requiring multiple cameras or special markers. Prior studies have demonstrated the feasibility of this approach [18,19]. Dunn et al. [18] validated a markerless single-camera system against marker-based three-dimensional motion analysis for walking and running, while Wang et al. [19] reported favorable agreement for most spatiotemporal gait parameters obtained from a single-camera computer-vision approach compared with marker-based multi-camera motion capture [18,19].
Although benefits exist, reducing multi-camera three-dimensional reconstruction to a single two-dimensional view introduces additional methodological challenges. Measurements may be affected by camera position and orientation, perspective disorientation, image-plane assumptions, landmark-detection error, and reduced ability to quantify movement occurring outside the calibrated plane [20]. Temporal measurement may also be influenced by video sampling frequency and the precision with which gait events, such as heel strike (initial contact of the heel with the ground) and toe-off (the point at which the toes leave the ground), can be identified [18,19].
Accordingly, the validity and test-retest reliability of spatiotemporal gait parameters derived from a conventional single-camera pose-estimation approach remain less established than those of multi-camera MMC systems. Establishing both properties is important because validity and reliability address different measurement characteristics. For example, a system may demonstrate systematic disagreement with a reference method, while producing repeatable measurements across sessions. Therefore, direct comparison with an established multi-camera MMC system can clarify which spatiotemporal parameters can be measured accurately and consistently using a simplified single-camera MMC configuration and which remain limited by the reduced spatial and temporal information available from a single video view.
Validation of gait-analysis technologies is commonly performed during standardized overground walking; however, everyday ambulation frequently occurs while individuals divide attention or carry external loads which may challenge an individual’s balance, coordination, or gait-adaptability. Such secondary demands may alter the spatiotemporal characteristics and may challenge pose-estimation and gait-event detection algorithms. Smartphone use during walking has been associated with changes in gait speed, cadence, and step and stride length [21], while external load carriage may similarly alter spatiotemporal gait characteristics [22]. Recent work has also extended markerless camera-based analysis to function gait tasks involving changes in walking speed, head movements, obstacle negotiation, and turning [23]. Accordingly, the present study included usual gait, backpack carriage, and phone use to determine whether single-camera measurement performance remained stable across conditions expected to alter gait behavior.
The purpose of the present study was to evaluate the validity (i.e., system agreement) and test-retest reliability of spatiotemporal gait parameter obtained using a single-camera pose-estimation MMC system, with an eight-camera artificial intelligence driven MMC system previously demonstrated to be valid and reliable for gait analysis serving as the reference system [12,14,15,17]. Although prior studies have established the feasibility and concurrent validity of single-camera MMC gait analysis, fewer studies have jointly characterized reference-system agreement and multi-session test-retest reliability of spatiotemporal outcomes obtained from a conventional RGB mobile-device workflow. The present study addresses this gap by comparing a single iPad camera with an eight-camera markerless reference system at Visit 1 and evaluating test-retest reliability of the single-camera MMC system across four visits and three common walking conditions that include walking challenges encountered in the real-world. We hypothesized that spatiotemporal parameters obtained from the single-camera MMC system would demonstrate agreement with the multi-camera reference system at Visit 1and test-retest reproducibility across four repeated testing sessions.
Materials and methods
Participants
The present study represents a secondary analysis of data collected as part of a larger study. An a priori sample-size analysis was conducted using G*Power (version 3.1.9.6; Heinrich Heine University Düsseldorf, Düsseldorf, Germany) before selection of the analytic cohort. Using a two-tailed bivariate correlation model, an anticipated correlation of 0.51, an alpha level of 0.05, and 90% statistical power, a minimum sample of 36 participants was required [24]. The anticipated effect was informed by previously published validation studies demonstrating moderate-to-strong relationships between markerless or single-camera motion-capture measurement and established reference methods [14,25,26].
Participant records from the parent study were screened in chronological order until 36 participants met the analytic eligibility criteria were identified. Eligibility required completion of the relevant testing protocol and sufficient usable single-camera and eight-camera markerless motion-capture data to contribute to the planned analyses. Thus, selection of the 36-participant analytic cohort (Mean: 18.7, SD: 1.7, range 18–27, 75% female, 30.6% racial minority, 33.3% Hispanic) was based on the a priori sample-size requirement followed by the chronological screening for data availability and analytic eligibility rather than post hoc selection based on study outcomes.
Procedures
The parent study used a repeated-measures design consisting of four visits separated by approximately two weeks. Participants walked back and forth along a 10-m walkway for one minute at a comfortable, self-selected pace. The 3-m capture region was positioned within the 10-m walkway, leaving approximately 2 m on one side and 5 m on the other; these regions provided acceleration and deceleration space dependent of the participant’s direction of travel.
Data within the 3-m capture region were collected simultaneously using an eight-camera MMC system (KinaTrax Inc., Boca Raton, FL, USA) and a single-camera MMC system consisting of an 11-inch iPad Pro with an M4 processor (Apple inc., Cupertino, CA, USA). The eight KinaTrax cameras, sampling at 100 Hz, were distributed around the 3-m capture space to provide multiple, synchronized viewing angles for each participant. As part of the standard laboratory data acquisition workflow, a static background image of the capture region was obtained for use in KinaTrax post-processing. The iPad, recording at 30 Hz using the native Camera application, was level-mounted with the camera at a height of 121 cm and positioned lateral to and approximately centered on the 3-m capture region, with the optical axis approximately perpendicular to the direction of gait progression, at a distance of 4 m from the walking path. The acquisition rate followed the parent study’s established recording protocol. The processing pipeline used the frame rate read from each video file rather than imposing a 30 Hz limit. Higher-frame-rate recordings were not evaluated in this secondary analysis. At 30 Hz, successive video frames were approximately 33.3 ms apart, compared with 10 ms for the 100 Hz reference system. This coarser temporal resolution may affect gait-event timing and the estimation of short gait phases, particularly swing duration.
Data were accessed between 04/20/2026 and 05/24/2026. The present analyses were restricted to three walking conditions: usual gait, bag-carrying, and phone-use. During the bag-carrying condition, participants wore a backpack containing an additional standardized 4.5 kg load. During the phone-use condition, participants held their personal mobile phone with both hands and were permitted to engage in a self-selected phone activity while walking. Other walking conditions included in the broader study protocol were not included in the present analysis. Participant depth relative to the camera was not quantified in the analysis and no sensitivity analysis was performed to estimate the effect of 5 − 10 cm depth deviations on spatial measurement. During each one-minute condition, participants walked continuously back and forth along the walkway, allowing multiple walking passes to be recorded. No familiarization trials were performed.
For the validity analysis, corresponding eight-camera and single-camera walking passes from Visit 1 were manually paired using synchronized trial timing and simultaneous visual inspection. Visual review was used to identify the heel strikes corresponding to those included in the reference-system gait calculation. Four heel strikes were selected for each retained validity trial and recorded in a pairing tracker by their sequence numbers. The single-camera pipeline automatically detected heel strikes and calculated their frame location, then retained the four selected events to compare corresponding gait cycles between systems. Manual selection did not replace the algorithmically determined heel-strike timings. The four selected events defined two complete gait cycles. Trials were excluded when the single-camera and reference-system recording could not be confidently paired, when fewer than four corresponding heel strikes could be identified, when the selected heel strikes did not yield two complete gait cycles for validity analysis, or when gait-event processing failed quality-control criteria described below. Therefore, the number of walking passes varied among participants; the requirement of two gait cycles referred to the number of cycles analyzed within each paired validity trial, not the number of walking passes collected during each one-minute condition. The analysis script retained the heel-strike numbers, heel-strike frame location, laterality, number of gait cycles, and an audit trail.
Data processing
Eight-camera MMC.
The eight-camera KinaTrax recording was processed using the laboratory’s established MMC workflow, consistent with previously published applications of the system [14,17]. Synchronized multi-camera video was processed by the KinaTrax system to generate three-dimensional markerless body trajectories, which were exported in C3D format. The resulting reference-system data were processed using Visual 3D (C-Motion, Inc., Germantown, MD, USA), from which trial-level spatiotemporal gait outputs were exported for analysis.
The Visual 3D output included gait speed, cycle time, stride length, and side-specific cadence, step length, stance duration, and swing duration. When bilateral values were available for cadence, step length, stance duration, and swing duration, left- and right-side values were averaged to generate the reference value used for comparison. Reference-system values were linked to the corresponding single-camera trial using the exact C3D filename retained during manual trial pairing. The analysis pipeline did not rematch trials statistically after processing; each iPad observation was lined to its predetermined reference-system walking pass.
Single-camera pose estimation and reconstruction.
Single camera recordings were processed in Python (version 3.11; Python Software Foundation, Wilmington, DE, USA) using OpenCV and MediaPipe Pose. MediaPipe Pose was implemented with (model_complexity = 2). Two-dimensional coordinates were extracted for the left and right heel, toe, and hip landmarks. Landmark observations with MediaPipe visibility values below 0.50 were treated as missing, and short gaps in landmark trajectories were interpolated before further processing.
To reduce high-frequency measurement noise, heel- and toe-position trajectories were filtered using a fourth-order Butterworth low-pass filter with a 4-Hz cutoff frequency. Walking direction was determined from the progression of the midpoint between the left and right hip landmarks. Heel and toe trajectories were then expressed relative to the hip midpoint and direction-normalized so that the same event detection procedure could be applied to walking passes recorded in either direction.
Pixel landmark coordinates were converted to planar real-world coordinates using a 0.77 x 0.77-m checkerboard positioned against the wall behind the participants within the single-camera field of view. The checkerboard was approximately 0.9 − 1.8 m behind the walking path. Four-fixed image coordinates representing the checkerboard were mapped to its known dimensions using OpenCV (cv2.getPerspectiveTransform). The resulting transformation was applied to heel coordinates, and stride length was calculated from horizontal displacement between successive same-limb heel strikes. No additional correction was applied for the depth difference between the checkerboard and walking path.
Candidate heel-strike events were identified from local maxima of the direction-normalized heel trajectory, while toe-off events were identified from local minima of the corresponding toe trajectory. Candidate events were required to be separated by at least 0.35 s. A velocity-based event-detection method was additionally calculated as a timing cross-check. When coordinate- and velocity-derived events occurred within three video frames, their frame location were averaged; otherwise, the coordinate-derived event was retained as the primary estimate. For validity trials, the automatically identified heel strikes were subsequently restricted to the four manually specified heel strikes recorded during the paired visual review. Toe-off events remained algorithmically determined.
A gait cycle was defined from one heel strike to the subsequent heel strike of the same limb. Within each cycle, cycle time was calculated as the elapsed time between consecutive same-limb heel strikes. Stance duration was calculated from the initial heel strike to the first ipsilateral toe-off occurring within the gait cycle, and swing duration was calculated from the toe-off to the subsequent ipsilateral heel strike. Cycles in which stance duration represented less than 45% or more than 85% of the complete gait cycle were rejected as implausible. Cadence was calculated as 120 divided by the mean cycle time.
Stride length was calculated from the absolute displacement along the walking axis between successive same-limb heel strikes after application of the planar perspective transformation. Projected stride lengths below 0.30 m or above 3.00 m were rejected by the processing pipeline. Mean stride length was calculated across retained cycles, step length was defined as one-half of mean stride length, and gait speed was calculated as mean stride length divided by mean cycle time.
For validity trials, the automatically detected heel-strike sequence was restricted to the four events selected during visual pairing. Selection was based on even sequence numbers; heel-strike frame locations remained algorithmically determined. Toe-off events also remained algorithmically determined. For reliability analyses, this manual event-selection restriction was not applied, and automatically detected events were used to construct all eligible gait cycles within each processed trial.
Statistical analysis
Statistical analyses were conducted in Python (version 3.11; Python Software Foundation, Wilmington, DE, USA) using pandas, NumPy, Pingouin, and Statsmodels. Statistical significance was set at α = 0.05. No adjustment for multiple comparisons was applied. P-values were interpreted as nominal, and findings were considered alongside effect estimates, confidence intervals, and agreement measures. Given the number of outcomes and comparisons, individual statistically significant findings should be interpreted cautiously. The distributions of inter-system differences were evaluated using Shapiro-Wilk tests and graphical agreement diagnostics. These tests assessed the unmodeled differences rather than residuals from the fitted mixed-effects models. Because each participant contributed multiple paired walking trials, statistical procedures accounting for the repeated-measures structure were used rather than treating individual trials as independent observations. Within-participant association between the single-camera and reference systems was assessed using repeated-measures correlation, with participant specified as the grouping variable. Before analysis, whole paired trials were excluded when iPad cadence was outside 60–180 steps·min ⁻ ¹, reference-system cadence was outside 60–180 steps·min ⁻ ¹, or iPad cycle time was outside 0.70–1.60 s. These thresholds were applied at the trial level, such that failure of a gait-event plausibility criterion resulted in removal of the complete paired trial from all spatiotemporal outcomes rather than removal of a single variable. Sensitivity analyses were performed to evaluate the influence of defined trial-level quality-control exclusions on validity and reliability estimates.
Validation analysis
Validity analyses were restricted to Visit 1 and included the usual-gait, bag, and phone walking conditions.
During source-level quality-control review, an additional criterion was introduced to identify clear reference-system gait-event processing failures. Left- and right-limb cycle duration were reconstructed from the sum of the corresponding Visual3D stance and swing durations. A trial was classified as internally inconsistent when at least one limb-specific cycle duration fell outside 0.70–1.60 s and the ratio of the longer to shorter bilateral cycle duration exceeded 1.50 s. This criterion was subsequently applied uniformly to all eligible paired trials. Analyses conducted before application of the numeric and internal reference-quality control rules were retained as sensitivity analyses.
Agreement was evaluated using repeated-measures Bland-Altman analysis. Differences scores were calculated as single-camera minus reference-system values, and overall bias was defined as the mean difference. Limits of agreement were calculated using a repeated-observations variance-component approach that incorporated both within-participant variance of the differences and between-participant heterogeneity in mean differences, while allowing unequal numbers of paired trials among participants. The 95% limits of agreement were calculated as:
Condition specific Bland-Altman estimates were additionally calculated for usual gait, bag, and phone conditions.
Linear mixed-effects models were used to evaluate proportional bias and determine whether inter-system differences varied by walking condition. For each spatiotemporal variable, the single-camera-minus-reference difference was modeled as the dependent variable. The centered mean of the two measurement systems and walking condition were entered as fixed effects, with participant included as a random intercept. A significant effect of measurement magnitude was interpreted as evidence of proportional bias. The overall effect of walking condition was assessed using a joint Wald test of the bag and phone coefficients, with usual gait serving as the reference condition. As an additional Bland-Altman variance diagnostic, Spearman correlation was calculated between each participant’s mean measurement magnitude and within-participant standard deviation of the inter-system differences.
Additional sensitivity analyses assessed the robustness of the findings to distributional departures and participant influence. Conditional mixed-model residuals were examined using quantile‒quartile, residual-versus-fitted, and absolute-residual-versus-fitted plots. A nonparametric participant level bootstrap used 2,000 resamples for each outcome. Participants were sampled with replacement, with all retained trials from each selected participant included together and repeated selections treated as separate participant clusters. The random-intercept mixed-effects models were refitted using restricted maximum likelihood, and repeated-measures correlations were recalculated in each resample. Percentile 95% confidence intervals were obtained from the bootstrap distributions without adjustment for multiple comparisons. Analyses were also repeated after omitting each participant in turn to assess sensitivity to individual participants.
Linear mixed-effects models were retained because they appropriately account for the dependence among repeated observations within participants, whereas the bootstrap analysis was performed as a sensitivity analysis to determine whether departures from normality materially influenced the fixed-effect estimates and confidence intervals.
Reliability analysis
Test-retest reliability of the single-camera system was evaluated separately for the usual-gait, bag, and phone conditions across the four visits, which were separated by approximately two weeks. The same eight-camera KinaTrax system used as the reference system in the present study has previously demonstrated high between-session reliability for spatiotemporal gait measures [17]; therefore, the present reliability analysis focused on the single-camera system.
Reliability analyses used the trial-level single-camera outputs. Trials were required to have successful processing status and at least two analyzable gait cycles. For each condition, retained trial values were averaged within participant and visit to produce one visit-level score for each spatiotemporal variable. Thus, the primary reliability statistics represent the reliability of a single visit and condition, rather than the reliability of an individual walking pass.
Participants with complete visit-level data across Visit 1 to Visit 4 for a given condition were included in that condition’s reliability analysis. The primary reliability estimate was the single-measure absolute-agreement intraclass correlation coefficient [ICC(A,1)]. ICC values were interpreted as poor (<0.50), moderate (0.50–0.75), good (0.75–0.90), and excellent (>0.90) [27]. Single-measure consistency ICCs [ICC(C,1)] and average-measure absolute-agreement ICCs [ICC(A,k)] were retained as secondary diagnostic estimates but were not used as the primary measure of test-retest reliability.
Measurement error was quantified using the standard error of measurement (SEM), calculated from the pooled standard deviation of the complete visit-level observation as:
The 95% minimal detectable change was calculated as:
A sensitivity analysis repeated the reliability calculation without the numeric cadence and cycle-time trial-level quality-control exclusions.
Results
For the validity analysis, 467 paired walking passes from Visit 1 met the initial trial-pairing and gait-cycle eligibility criteria across the usual-gait, bag, and phone conditions. Nine paired trials were subsequently excluded by the final whole-trial quality-control criteria because of implausible single-camera gait-event estimates, implausible reference-system cadence values, and/or marked internal inconsistency in bilateral reference-system gait-event timing. The final validity dataset consisted of 458 paired trials (167 usual-gait trials, 109 bag-condition trials 182 phone-condition trials). Of note, two participants did not provide qualifying trials to the bag condition. Across the three conditions combined, participants contributed a mean of 12.7 retained paired trials (range, 4–18). Thus, the paired observations were derived from three separate one-minute walking conditions rather than from a single one-minute walking session.
Shapiro-Wilk tests indicated statistically significant deviations from normality for the inter-system difference distributions of all seven spatiotemporal variables (W = 0.797‒0.959; all p < 0.001). Repeated-measures Bland-Altman variance diagnostics showed no significant associate between participant-level measurement magnitude and within-participant variability in the differences for cadence, gait speed, cycle time, stance duration, or swing duration (all p > 0.21). A modest positive association was observed for both step length and stride length (Spearman ρ = 0.331, p = 0.049), indicating greater within-participant variability at larger spatial measurement.
Conditional residual diagnostics showed departures from normality, including skewness and heavy tails, particularly for temporal outcomes. All 2,000 participant-level bootstrap resamples were successfully fitted for each outcome (Table 1). Bootstrap 95% confidence intervals excluded zero for all seven proportional-bias slopes and for six repeated-measures correlations; the interval for swing-duration correlation interval included zero (−0.040‒0.112). Most individual walking-conditions contrasts yielded the same conclusions regarding inclusion of zero as the original model-based intervals. An exception was the bag-versus-usual-gait contrast for cycle time, whose bootstrap interval excluded zero (0.0011‒00155 s), although its original model-based interval included zero. Leave-one-participant-out analyses retained positive proportional-bias slopes for all outcomes, although their magnitudes varied, particularly for stance duration.
Validity
Repeated-measures correlation demonstrated statistically significant positive within-participant associations between the single-camera and eight-camera MMC systems for six of the seven spatiotemporal outcomes (Fig 1). Gait speed demonstrated the strongest association, followed by stance duration, cycle time, and cadence. Step and stride length demonstrated weaker association, where swing duration was not significantly associated between systems.
Diamonds represent repeated-measures correlation coefficients (rrm) and horizontal error bars represent 95% confidence intervals. Filled diamonds indicate statistically significant associations (p < 0.05), whereas the open diamond indicates a nonsignificant association. The vertical dashed line denotes rrm = indicating no within-participant linear association between measurement systems.
Despite small mean system differences, repeated-measures limits of agreement varied across outcomes (Table 2; Fig 2). Positive differences indicate higher single-camera estimates and negative differences indicate lower single-camera estimates and negative differences indicate lower single-camera estimates relative to the eight-camera reference system.
Differences were calculated as single-camera minus eight-camera markerless motion capture values and are plotted against the mean of the two systems. Points represent individual paired walking trials and are differentiated by usual-gait, bag, and phone conditions. Dashed lines indicate the overall repeated-measures mean bias and 95% limits of agreement. A: cadence, B: gait speed, C: cycle time, D: stance duration, E: swing duration, F: step length, G: stride length.
Significant proportional bias was observed for all seven spatiotemporal outcomes (all p < 0.001), indicating that inter-system differences varied with measurement magnitude. Walking conditional additionally influenced inter system bias for cadence, gait speed, cycle time, step length, and stride length, but not stance or swing duration (Table 3).
Reliability
Following participant, visit, condition, processing-status, and gait-cycle eligibility screening, 2,916 single-camera trials were available for the reliability analysis before numeric trial-level quality control. Twenty-seven trials were excluded because cadence and/or cycle-time values fell outside the established gait-event plausibility criteria, leaving 2,889 trials across four visits and the usual-gait, bag, and phone conditions. Thirty-three of the 36 participants had complete visit-level data across all four visits and were included in each condition-specific test-retest reliability analysis.
Primary single-measure absolute-agreement reliability ICC(A,1), ranged from 0.499 to 0.862 across variables and walking conditions (Table 3). Gait speed demonstrated good reliability across all three walking conditions, whereas cadence cycle time, and stance duration demonstrate moderate reliability during usual gait and the bag condition and good reliability during the phone condition. Swing duration demonstrated the lowest reliability, ranging from poor to moderate across conditions (0.499–0.562 across conditions). Measurement-error estimates varied by parameter but were generally similar across walking conditions (Table 4).
Discussion
The purpose of this study was to evaluate the validity and test-retest reliability of spatiotemporal gait parameters obtained from a single-camera MMC system using an eight-camera MMC system as the reference. Overall, the findings indicate that performance of the single-camera approach was parameter dependent rather than uniformly valid or reliable. Gait speed demonstrated the strongest within-participant association with the reference system, followed by stance duration and cycle time, whereas cadence showed a weaker association and swing duration showed essential no within-participant associations. Despite these differences in association, mean inter-system biases were relatively small for most variables. However, mixed-effects analyses identified significant proportional bias for every spatiotemporal parameter, indicating that inter-system differences changed across the measurement range. Test-retest reliability of the single-camera MMC system similarly varied by parameter and walking condition, ranging from poor to good using the primary single-measure absolute agreement ICC. These findings provide a more conservative assessment of validity than the preliminary analysis because the final statistical approach accounted for multiple paired walking trials contributed by each participant within Visit 1. Previous validation studies of multi-camera MMC have reported strong agreement for several spatiotemporal gait variables. Ripic et al. [14], for example, reported strong agreement between multi-camera MMC and reference methods for gait speed, stride length, step length, cycle time, and step time, while stance- and swing-related measures were more sensitive to differences in gait-event detection methods. The current results are directionally consistent with those findings for gait speed and cycle time, but agreement was weaker for several outcomes when movement was estimated from a single 30-Hz camera rather than a synchronized multi-view three-dimensional reconstruction. Importantly, the repeated-measures correlation coefficients reported here should not be directly compared numerically with Pearson correlations, ICCs, or concordance coefficients reported in previous validation studies because these statistics quantify different aspects of association or agreement.
Gait speed demonstrated the strongest within-participant association between systems (rrm = 0.588) and a small mean bias of −0.019 m·s−1, although the repeated-measures limits of agreement extended from −0.195 to 0.157 m·s−1. Cycle time and stance duration also demonstrated moderate within-participant associations, with mean biases of −0.025 s and −0.042 s, respectively. These findings are broadly consistent with previous work demonstrating favorable performance of multi-camera MMC for gait speed and cycle-related measures [12,14]. In stroke survivors, for example, the same KinaTrax multi-camera platform has demonstrated good-to-excellent validity for gait speed, stride length, and several temporal outcomes [12]. Nevertheless, the significant proportional bias identified for all outcomes in the present study indicates that mean bias alone should not be interpreted as evidence that the single-camera and multi-camera systems are interchangeable. Mean bias represents the average difference across the sample; individual differences between systems may be considerably larger and may change with measurement magnitude.
Swing duration provides an especially clear example of the distinction between mean bias and association. The mean bias for swing duration was only + 0.017s, yet its repeated-measures correlation was essentially zero (rrm = 0.026, p = 0.599). Thus, the two methods produced similar values on average while failing to consistently track within-participant changes in swing duration across trials. This pattern is plausible given the manner in which gait events were reconstructed by the single-camera MMC system. Heel strikes used for validity were manually matched to the corresponding reference-system gait cycles, whereas toe-off remained algorithmically detected from the two-dimensional toe trajectory. Swing duration therefore depended directly on the timing of an automatically detected toe-off and the subsequent heel strike. At the single-camera sampling rate of 30 Hz, adjacent frames are separated by approximately 0.033s, making relatively short phase-duration measures particularly sensitive to frame-level event-timing differences. Prior multi-camera validation work has similarly identified stance and swing measures as sensitive to gait-event definition and detection methodology [14].
Cadence also remained less robust than gait speed and several other temporal measures. The single-camera MMC system overestimated cadence by an average of 2.81 steps·min−1, with repeated measures limits of agreement from −9.46 to 15.08 steps·min−1 and a repeated -measures correlation of 0.446. The mixed-effects model additionally demonstrated proportional bias, indicating that cadence error was dependent on measurement magnitude. Because cadence in the single-camera pipeline was calculated from mean gait-cycle duration, inaccuracies of only a few video frames in heel-strike timing can propagate into the cadence estimate. The sensitivity analysis further demonstrated the importance of gait-event quality control: when trials containing clear gait-event processing failures were retained, the repeated-measures correlation for cadence decreased to approximately zero and its limits of agreement became extremely wide. This finding indicates that cadence is particularly vulnerable to isolated gait-event processing failures and emphasizes the importance of trial-level quality-control procedures when validating automated gait-analysis systems. The condition effects modeled here represented differences between systems rather than direct estimates of biological changes in cadence. The analysis did not separately quantify heel-landmark occlusion, tracking lag, or event-timing error by condition. Therefore, the observed condition-dependent disagreement cannot be attributed specifically to occlusion or to genuine changes in step frequency.
The quality-control findings also demonstrate that the eight-camera MMC reference should not be regarded as error-free. Source-level review identified a small number of trials in which reference-system gait-event outputs were internally inconsistent between limbs. Application of the same objective quality-control rule across all eligible trials removed these processing failures before primary analysis. This distinction is important because disagreement between two measurement systems cannot automatically be attributed to the single-camera method. Prior work has established strong validity and reliability of the multi -camera KinaTrax measurements, but any MMC system remains dependent on successful pose reconstruction and gait-event identification [12,14,17]. The substantial deterioration observed in the sensitivity analysis when failed trials were retained illustrates how a small number of event-detection errors can disproportionately influence agreement estimates, particularly for temporally derived outcomes.
Spatial measures showed a somewhat different pattern. Step length and stride length demonstrated relatively small mean biases of −0.026 m and −0.051 m, respectively, but only modest repeated-measures associations with the reference system (rrm ≈ 0.33). Moreover, the mixed-effects models demonstrated proportional bias and significant effects of walking condition for both variables. The single-camera pipeline converted two-dimensional heel coordinates into metric coordinates using a fixed 0.77 x 0.77-m checkerboard reference and a planar perspective transformation. This approach permits spatial measurements using a single conventional camera but does not reproduce the three-dimensional multi-view reconstruction available from the eight-camera reference system. These limitations are consistent with previous single-camera work demonstrating that gait-estimation accuracy is sensitive to camera perspective, viewing angle, and distance from the participant [18,19]. Consequently, perspective effects, pose-estimation error, and movement outside the calibrated two-dimensional plane may contribute to disagreements in spatial estimates. The checkerboard and walking path were at different distances from the camera, and the spatial conversion did not correct for this separation. The checkerboard-derived scale therefore may not accurately represent distances in the participants’ walking plane. This limitation affects interpretation of stride length, derived step length, and gait speed. The contribution of calibration depth to the observed inter-system differences was not separately quantified and cannot be distinguished from landmark-tracking and gait-event detection errors in the present analysis. These findings contrast with previous multi-camera MMC studies reporting strong agreement for stride and step length, highlighting an important tradeoff between the accessibility of a single-camera approach and the additional spatial information available from synchronized multi-view systems [12,14].
An additional consideration is that step length in the present single-camera processing pipeline was calculated as one-half of mean stride length rather than from the distance between successive contralateral heel strikes. Therefore, step and stride length are mathematically related within the single-camera measurements and should not be interpreted as fully independent validation outcomes. This approach also cannot characterize left-right step-length asymmetry. The nearly identical validity associations and identical reliability coefficients observed for step and stride length are consequently expected from the processing method. Future implementations should calculate individual left and right step length directly from contralateral heel-strike positions, particularly if the system is to be applied to populations in which gait asymmetry is clinically relevant.
Test-retest reliability was lower and more variable than suggested by the preliminary ICC(3,k) analysis. Using ICC(A,1), which quantified the absolute agreement of a single visit-level score, gait speed demonstrated good reliability in all three walking conditions (ICC = 0.774–0.804). Cadence, cycle time, stance duration, step length, and stride length generally demonstrated moderate-to-good reliability depending on condition, whereas swing duration demonstrated the lowest reliability (ICC = 0.499–0.562). Step and stride length demonstrated their highest reliability during the phone condition (ICC = 0.862). These results indicate that reliability should be described as variable and condition dependent, rather than as uniformly good to excellent.
The lower ICC estimates relative to previous MMC reliability studies should be interpreted in light of important methodological differences. Schoenwether et al. [17] reported predominantly excellent between-session reliability for spatiotemporal parameters obtained using the eight-camera KinaTrax system, with between-session ICCs exceeding 0.90 for most outcomes. However, that study evaluated multi-camera MMC and used an average-measures ICC(3,k), whereas the primary analysis in the present study used a single-measure absolute-agreement ICC(A,1) [17]. Average-measures ICCs generally represent the reliability of an average of repeated measurements and should not be directly compared with the reliability of a single visit-level score. The more conservative ICC formulation used here therefore provides a more appropriate estimate of how reproducible one measurement session from the single-camera system would be in practice.
Absolute measurement-error estimations provide additional context. During usual gait, cadence demonstrated an SEM of 4.07 steps·min−1 and an MDC95 of 11.28 steps·min−1; corresponding values across the bag and phone conditions were similar. These values are reasonably close to previous multi-camera MMC results. Schoenwether et al. reported a between-session cadence of SEM of 3.593 steps·min−1 and MDC of 9.958 steps·min−1 for the least reliable cadence measure in their analysis [17]. Similarly, Riazati et al. reported low absolute measurement error for most spatiotemporal parameters obtained using multi-camera Theia3D, including an SEM of 2.92 steps·min−1 for self-selected cadence and 0.07 m·s−1 for gait speed [16]. In the present study, gait-speed SEM values ranged from approximately 0.051 to 0.054 m·s−1 across conditions. Thus, although relative reliability as quantified by ICC(A,1) was lower than reported for some multi-camera systems, the absolute measurement error for several variables remained of a similar order of magnitude to previous MMC studies.
The distinction between relative and absolute reliability is particularly relevant for swing duration. Swing duration demonstrated poor-to-moderate ICC(A,1) values but relatively small SEM values of approximately 0.014–0.016 s and MDC95 values of approximately 0.039–0.043 s. A lower ICC does not necessarily indicate large absolute measurement error because ICC magnitude is influenced by between-participant variability. The combination of a low ICC and small SEM suggest that swing-duration values occupied a relatively restricted range in the healthy sample while still showing insufficient stability to preserve between-participant differences across repeated visits. For applications involving longitudinal monitoring, the MDC95 values may therefore be more directly useful: observed changes smaller than the MDC95 cannot be confidently distinguished from expected measurement variability.
Walking condition also influenced inter-system differences for cadence, gait speed, cycle time, step length, and stride length. Step and stride length nevertheless demonstrated their highest test-retest ICCs during the phone condition. These findings are compatible because the analyses addressed different comparisons: reliability assessed agreement among repeated single-camera visit-level measurement, whereas validity assessed agreement between the single-camera measurements can retain a systematic difference from the reference system. Furthermore, the joint condition test assessed whether the bag and phone coefficients were jointly zero after accounting for measurement magnitude; it did not directly identify which condition had the greatest absolute measurement error. Condition-specific reliability should therefore not be interpreted as evidence of interchangeability between systems.
Several characteristics of the single-camera processing pipeline may have contributed to the observed results. The use of a fixed camera configuration, a known checkerboard spatial reference, and planar perspective transformation enabled metric spatial estimates from two-dimensional video. A fourth-order Butterworth low-pass filter with a 4-Hz cutoff was used to reduce high-frequency noise in landmark trajectories, and gait events were estimated from direction-normalized heel and toe trajectories. These design choices provide a comparatively simple processing approach that can operate from conventional RGB video; however, they also introduce limitations that are not present to the same extent in multi-camera three-dimensional systems. The 30-Hz frame rate limits temporal resolution, the two-dimensional reconstruction cannot directly quantify depth, and event timing remains dependent on successful MediaPipe landmark tracking. The validity analysis required manual pairing of walking passes and selection of corresponding heel strikes from the automatically detected event sequence. Although heel-strike and toe-off timings remained algorithmically determined, this manual selection means that the reported validity represents a researcher-assisted comparison rather than a fully automated workflow. The reliability analysis did not apply this manual event-selection restriction. Nevertheless, repeatability without manual event selection does not establish agreement with the reference system under fully automated conditions. Further validation without manual event selection is needed before the reported validity can be extended to an unattended clinical workflow. Further development should evaluate automated identification of the analysis interval and eligible gait cycles, together with automated quality-control flags. Agreement with a reference system should then be reassessed without manual event selection.
The controlled laboratory environment also limits generalizability. Participants were young, medically healthy adult, camera position and viewing geometry were fixed, and the walking path was standardized. The system’s performance may differ in older adults or individuals with neurological, musculoskeletal, or other gait disorders, who may demonstrate greater asymmetry, altered foot trajectories, use of assistive devices, or greater variability in gait-event timing. Previous multi-camera MMC research has demonstrated the feasibility of markerless spatiotemporal gait measurement in both healthy and clinical populations, but those findings cannot be assumed to transfer directly to the present single-camera reconstruction [12,14,16,17]. Tracking performance was not evaluated separately by clothing characteristics or body mass index. Although the processing pipeline applied a landmark-visibility threshold, this does not establish comparable performance across body types, clothing, or recording environments. Performance under variable lighting and occlusion from assistive devices remains to be established. Multiple outcomes and walking-condition effects were tested without adjustment for multiple comparisons, increasing the possibility of false-positive findings. Individual significant associations and condition effects therefore require cautious interpretation, particularly when p-values are close to the nominal significance threshold.
Residual diagnostics indicated that normality and constant-variance assumptions were not uniformly supported. Participant-level bootstrap analyses nevertheless supported the principal findings of proportional bias across all outcomes and positive within-participant associations for six outcomes, with no clear association for swing duration. The differing interval conclusions for the bag-versus-usual-gait cycle-time contrast indicate some secondary findings are sensitive to the method used to estimate uncertainty. These analyses therefore support the main interpretation while warranting caution regarding individual condition effects and the precision of model-based estimates. The bootstrap analyses assessed regression coefficients and correlation; they did not establish the empirical coverage of the reported normal-theory limits of agreement.
Taken together, the present findings support the feasibility, rather than complete interchangeability, of single-camera MMC for selected spatiotemporal gait outcomes. Gait speed demonstrated the most favorable combination of within-participant association, small systemic bias, and test-retest reliability. Cycle time and stance duration also showed useful associations with the reference system, whereas cadence demonstrated greater magnitude-dependent error and swing duration showed little within-participant association between systems. Spatial measurement exhibited small average biases but only modest associations and significant proportional and condition-dependent bias. These results demonstrate that an accessible single-camera approach can provide reproducible information for several gait parameters under controlled conditions, but parameter-specific measurement of error and systematic differences from multi-camera MMC must be considered when interpreting individual measurements or longitudinal change.
Conclusion
The single-camera MMC system demonstrated parameter-dependent validity and test-retest reliability for spatiotemporal gait assessment under controlled laboratory conditions. Gait speed showed the strongest within-participant association with the eight-camera MMC reference system and demonstrated good test-retest reliability across walking conditions. Cycle time and stance duration also demonstrated moderate associations with the reference system, whereas cadence showed greater variability and swing duration demonstrated little within-participant association between systems. Step and stride length exhibited small mean biases but comparatively modest associations with the reference system. Because single-camera step length was calculated as half of stride length, these findings do not establish the validity of independently measured step length or the ability to assess left−right step-length asymmetry. Significant proportional bias across all spatiotemporal variables further indicates that the two systems should not be considered interchangeable across the observed measurement ranges.
Test-retest reliability of the single-camera MMC system ranged from poor to good and varied by both gait parameter and walking condition. These findings support the feasibility of single camera MMC for measuring selected spatiotemporal gait parameters, particularly gait speed, while highlighting limitations associated with temporal gait-event detection and two-dimensional spatial reconstruction. These findings are limited to a predominantly female sample of young, healthy adults assessed under controlled laboratory conditions and should not be generalized to older adults or individuals with gait disorders. Validation in those populations, including individuals with gait asymmetry or assistive-device use, and in less controlled environments is required before clinical application.
Acknowledgments
The authors thank the study participants for their time and participation. The authors also acknowledge the research interns of the University of Miami Sports Medicine and Motion Analysis Laboratory and the Bridging Research on Anxiety, Innovation, and Neuroscience (BRAIN) Group who assisted with data collection throughout the study. The authors further thank Veeraj Jain for formatting the manuscript in accordance with PLOS ONE submission requirements.
References
- 1. Baker R, Esquenazi A, Benedetti MG, Desloovere K. Gait analysis: clinical facts. Eur J Phys Rehabil Med. 2016;52(4):560–74. pmid:27618499
- 2. Wren TAL, Tucker CA, Rethlefsen SA, Gorton GE 3rd, Õunpuu S. Clinical efficacy of instrumented gait analysis: Systematic review 2020 update. Gait Posture. 2020;80:274–9. pmid:32563727
- 3. Baker R. Gait analysis methods in rehabilitation. J Neuroeng Rehabil. 2006;3:4. pmid:16512912
- 4. Hulleck AA, Menoth Mohan D, Abdallah N, El Rich M, Khalaf K. Present and future of gait assessment in clinical practice: Towards the application of novel trends and technologies. Front Med Technol. 2022;4:901331. pmid:36590154
- 5. Whittle MW. Clinical gait analysis: A review. Hum Mov Sci. 1996;15(3):369–87.
- 6. Elbaz A, Mor A, Segal G, Debi R, Shazar N, Herman A. Novel classification of knee osteoarthritis severity based on spatiotemporal gait analysis. Osteoarthritis Cartilage. 2014;22(3):457–63. pmid:24418677
- 7. Givon U, Zeilig G, Achiron A. Gait analysis in multiple sclerosis: characterization of temporal-spatial parameters using GAITRite functional ambulation system. Gait Posture. 2009;29(1):138–42. pmid:18951800
- 8. Lemke MR, Wendorff T, Mieth B, Buhl K, Linnemann M. Spatiotemporal gait patterns during over ground locomotion in major depression compared with healthy controls. J Psychiatr Res. 2000;34(4–5):277–83. pmid:11104839
- 9. Stout A, Alvarez M, Fisanick L, Kauffman M, Dawood M, Mabbun K, et al. Feasibility of machine learning classification of depression and anxiety symptoms among college students using 3D gait and sit-to-walk biomechanics. Gait Posture. 2026;126:110123. pmid:41687574
- 10. Scataglini S, Abts E, Van Bocxlaer C, Van den Bussche M, Meletani S, Truijen S. Accuracy, Validity, and Reliability of Markerless Camera-Based 3D Motion Capture Systems versus Marker-Based 3D Motion Capture Systems in Gait Analysis: A Systematic Review and Meta-Analysis. Sensors (Basel). 2024;24(11):3686. pmid:38894476
- 11. Varcin F, Boocock MG. The accuracy, validity and reliability of Theia3D markerless motion capture for studying the biomechanics of human movement: A systematic review. Artif Intell Med. 2026;173:103332. pmid:41518893
- 12. Alammari BJ, Schoenwether B, Ripic Z, Kirk-Sanchez N, Eltoukhy M, Bishop L. Validity of AI-Driven Markerless Motion Capture for Spatiotemporal Gait Analysis in Stroke Survivors. Sensors (Basel). 2025;25(17):5315. pmid:40942744
- 13. Ripic Z, Nienhuis M, Signorile JF, Best TM, Jacobs KA, Eltoukhy M. A comparison of three-dimensional kinematics between markerless and marker-based motion capture in overground gait. J Biomech. 2023;159:111793. pmid:37725886
- 14. Ripic Z, Signorile JF, Kuenze C, Eltoukhy M. Concurrent validity of artificial intelligence-based markerless motion capture for over-ground gait analysis: A study of spatiotemporal parameters. J Biomech. 2022;143:111278. pmid:36063770
- 15. Ripic Z, Signorile JF, Best TM, Jacobs KA, Nienhuis M, Whitelaw C, et al. Validity of artificial intelligence-based markerless motion capture system for clinical gait analysis: Spatiotemporal results in healthy adults and adults with Parkinson’s disease. J Biomech. 2023;155:111645. pmid:37216895
- 16. Riazati S, McGuirk TE, Perry ES, Sihanath WB, Patten C. Absolute Reliability of Gait Parameters Acquired With Markerless Motion Capture in Living Domains. Front Hum Neurosci. 2022;16:867474. pmid:35782037
- 17. Schoenwether B, Ripic Z, Nienhuis M, Signorile JF, Best TM, Eltoukhy M. Reliability of artificial intelligence-driven markerless motion capture in gait analyses of healthy adults. PLoS One. 2025;20(1):e0316119. pmid:39841651
- 18. Dunn M, Haake S, Wheat J, Goodwill S. Validation of a Single Camera, Spatio-temporal Gait Analysis System. Procedia Eng. 2014;72:243–8.
- 19. Wang H, Su B, Lu L, Jung S, Qing L, Xie Z, et al. Markerless gait analysis through a single camera and computer vision. J Biomech. 2024;165:112027. pmid:38430608
- 20. Needham L, Evans M, Cosker DP, Wade L, McGuigan PM, Bilzon JL, et al. The accuracy of several pose estimation methods for 3D joint centre localisation. Sci Rep. 2021;11(1):20673. pmid:34667207
- 21. Bruyneel AV, Reinmann A, Gafner SC, Sandoz JD, Duclos NC. Does texting while walking affect spatiotemporal gait parameters in healthy adults, older people, and persons with motor or cognitive disorders? A systematic review and meta-analysis. Gait Posture. 2023;100:284–301.
- 22. Talarico MK, Haynes CA, Douglas JS, Collazo J. Spatiotemporal and kinematic changes in gait while carrying an energy harvesting assault pack system. J Biomech. 2018;74:143–9. pmid:29752054
- 23. Darici O, Cabak C, Wong JD. A low-cost markerless motion capture system to automate functional gait assessment: Feasibility study. PLoS One. 2026;21(6):e0346606. pmid:42228716
- 24. Faul F, Erdfelder E, Buchner A, Lang A-G. Statistical power analyses using G*Power 3.1: tests for correlation and regression analyses. Behav Res Methods. 2009;41(4):1149–60. pmid:19897823
- 25. Clark RA, Bower KJ, Mentiplay BF, Paterson K, Pua Y-H. Concurrent validity of the Microsoft Kinect for assessment of spatiotemporal gait variables. J Biomech. 2013;46(15):2722–5. pmid:24016679
- 26. Pottorf O, Vapne D, Ghigiarelli J, Haase K. Reliability and Concurrent Validity of a Markerless, Single Camera, Portable 3D Motion Capture System for Assessment of Glenohumeral Mobility. Int J Sports Phys Ther. 2023;18(5):1176–85. pmid:37795320
- 27. Koo TK, Li MY. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med. 2016;15(2):155–63. pmid:27330520