Introduction

Accurate assessment of gait is pivotal for diagnosing and managing a wide range of neurological, musculoskeletal, and vestibular disorders, including Parkinson’s disease, cognitive impairments, spinal cord injury, stroke, and dizziness1,2,3,4,5,6. Gait impairments, such as reduced stride length, prolonged double support time, and diminished gait speed, can serve as indicators of pathology or markers of increased fall risk and mobility decline7,8. Yet despite its clinical value, gait assessment remains underutilized in routine clinical care and when performed, it is typically limited to subjective observation and the use of standardized rating scales9,10. However, the above approaches require specialized training and may lack the sensitivity needed to detect subtle or progressive changes associated with disease progression or therapeutic intervention.

Digital health technologies, including motion capture systems, wearable sensors, and multi-camera markerless systems, have facilitated objective gait analysis and enable precise quantification of spatiotemporal gait parameters that can help detect and manage gait and mobility impairments11. However, these systems are often expensive, require specialized devices, controlled environments, and technical expertise, making them impractical for widespread use in primary care, community-based, or home-based settings. Even recently introduced low-cost, two-camera solutions, such as OpenCap, still require deliberate camera placement, calibration procedures, and sufficient space to maintain a clear field of view, which limits usability outside of controlled recording environments12.

Recent advances in computer vision and machine learning have enabled video-based gait analysis using a single camera, offering a potentially scalable solution for gait analysis6,13,14,15,16. However, most existing methods rely on sagittal-view recordings, requiring open, unobstructed walkways with sufficient distance between the camera and the participant, limiting feasibility in typical clinical environments. Because only one limb is fully visible at any time, sagittal-view methods may provide incomplete information and fail to capture mediolateral instability or frontal-plane gait deviations that might be clinically meaningful and could be captured with frontal-view videos17. Moreover, few single-camera methods have been validated against reference measures, across diverse patient populations, or in real-world settings.

The objectives of this study were to introduce, validate and assess the clinical utility of VisionMD-Gait, a scalable, user-friendly, and open-source platform designed to estimate spatial, temporal, and spatiotemporal gait parameters from frontal-view videos18. Unlike other video-based gait analysis tools, VisionMD-Gait is fully open source and free to use, easily customizable, and requires no coding knowledge, allowing clinicians, therapists, and researchers with minimal technical expertise to incorporate it effectively into clinical practice or research workflows. Furthermore, all data are processed locally on the user’s device, maximizing patient privacy and data protection by ensuring no clinical data leave the computer on which they are stored.

VisionMD-Gait leverages recent advances in 3D pose estimation and video-based reconstruction of gait kinematics to compute clinically relevant gait parameters from frontal-view smartphone recorded videos15,19. Fig. 1 shows an overview of the VisionMD-Gait pipeline which includes three primary components: (i) Subject Selection, (ii) Segment Selection, and (iii) Analysis. In the Subject Selection component, users upload a gait video to VisionMD-Gait and the software automatically localizes the position of the subject of interest in the video. In the Segment Selection component, users define the start and end times of the different gait segments in the video using frame-level precision. Finally, in the Analysis component, VisionMD-Gait uses machine learning algorithms for 3D pose and gait kinematics reconstruction to track body movements during gait and compute quantitative gait parameters that help characterize any potential gait impairments.

Fig. 1
figure 1

Overview of the VisionMD-Gait processing pipeline. VisionMD-Gait integrates state-of-the-art monocular video analysis methods into an open-source platform for clinical gait assessment. (i) Subject Selection: VisionMD-Gait automatically detects and tracks the subject of interest in a frontal-view video. (ii) Segment Selection: Users define gait segments with frame-level precision. (iii) Analysis: The system reconstructs 3D kinematics and computes spatiotemporal gait parameters including gait speed, step length, cadence, stance time, swing time, and double support time. Kinematics signals and gait parameters can be stored for future use in external tools.

Full size image

VisionMD-Gait was designed for seamless integration into standard clinical environments, requiring only a smartphone camera and a hallway or corridor that allows participants to walk unperturbed along a straight path. We validated VisionMD-Gait against a research-grade wearable motion capture system both healthy adults and individuals experiencing dizziness. We hypothesize that (i) video-derived gait parameters would strongly correlate with sensor-derived measures in both healthy controls and dizziness groups, and (ii) VisionMD-Gait derived gait metrics would significantly differentiate healthy controls from people with dizziness.

VisionMD-Gait is an open-source tool that represents an accessible, scalable solution for gait assessment in clinical and research settings and complements subjective observations and standardized rating scales. The software tool is available for Windows, macOS, and Linux (www.VisionMD.ai). This work contributes to the growing field of digital health technologies for mobility and demonstrates the feasibility of low-cost, high-impact tools for routine gait monitoring.

Results

Participants and data

To validate the output of VisionMD-Gait, we performed a validation study comparing video-derived gait parameters provided by VisionMD-Gait against a research grade wearable 3D motion capture system (Noraxon Ultium Motion) in a diverse population that included 24 healthy adults (13 females, 11 males; age: 51.96 ± 24.23 years) and 10 adults with dizziness receiving ambulatory clinical care at a vestibular clinic (8 females, 2 males, age: 51.73 ± 24.77 years).

After removing outliers, we analyzed a total of 269 laps in the 8-meter walkway (187 laps completed by healthy controls and 77 laps completed by persons with dizziness). All participants completed at least 4 laps, while some participants completed up to 10 laps.

Validation

Across both healthy controls and individuals with dizziness, video-derived gait parameters from VisionMD-Gait showed strong agreement with reference measures obtained from wearable sensors. Tables 1 and 2 summarize sensor- and video-based estimates (mean ± SD) obtained for each group, averaged across the left and right legs, along with the corresponding MAE%, and Pearson’s correlation coefficients. As shown in these tables, correlations were consistently high for all gait parameters (r = 0.76–1.00 for healthy controls; r = 0.70–1.00 for dizziness group, all p < < 0.001 after Bonferroni correction), indicating that video-derived estimates closely tracked sensor-based measures. Mean absolute errors were generally low across parameters and populations, with step duration, stride duration, and cadence showing MAE% values below 1%. Larger relative errors were observed for double support time, step length, and gait speed, although all errors remained below 10%.

Table 1 Agreement between video-based and sensor-based gait parameters for healthy controls.
Full size table
Table 2 Agreement between video-based and sensor-based gait parameters for individuals with dizziness.
Full size table

To further characterize bilateral gait dynamics, we examined left and right limb parameters independently (Tables S1 and S2). This supplementary analysis revealed patterns consistent with those obtained from averaged results, with strong correlation and low MAE% across most measures. Notably, left and right step length exhibited relatively larger errors, with MAE% around 10%, and lower correlations (r = 0.59–0.68 for both groups, all p < < 0.001 after Bonferroni correction).

Figure S1 and Figure S2 (included in the supplementary material) show the results from the Bland-Altman analysis for each group, further confirming the high level of agreement between sensor- and video-based estimates. Across all temporal and spatiotemporal parameters, mean differences (bias) were close to zero, and the 95% limits of agreement were narrow and symmetric, indicating no systematic over- or under-estimation across the range of gait behaviors. Spatial parameters—particularly step length—exhibited slightly wider limits of agreement, in line with the higher MAE% and lower correlation results. Importantly, no evidence of proportional bias was observed, and the overall agreement profiles were comparable between healthy controls and individuals with vestibular dizziness.

Clinical application

Table 3 compares the video-derived gait parameters between healthy controls and individuals with dizziness. The table presents estimates for each group (mean ± SD), the Linear Mixed Model coefficient associated with the Group variable (95% CI) and p-values (after Bonferroni correction). As shown in the table, there were significant differences in almost all gait measures despite the absence of clinically overt gait impairments. Compared to healthy controls, the dizziness group exhibited reduced gait speed (1.29 ± 0.17 m/s vs. 1.05 ± 0.11 m/s, p = 0.004) and cadence (108.75 ± 6.45 steps/min vs. 100.66 ± 7.81 steps/min, p = 0.013), alongside longer step duration (0.55 ± 0.04 s vs. 0.60 ± 0.04 s, p = 0.011), stride duration (1.11 ± 0.07s vs. 1.20 ± 0.09 s, p = 0.011), stance time (0.70 ± 0.05 s vs. 0.76 ± 0.06 s, p = 0.011), swing time (0.41 ± 0.02 s vs. 0.44 ± 0.03 s, p = 0.012), and double support time (0.29 ± 0.02 s vs. 0.32 ± 0.03 s, p = 0.016). In all cases, the residuals passed the test for normality, homoscedasticity, and independence, validating the accuracy of our results  (see Figure S3 in supplementary material).

Table 3 Comparison of video-derived gait parameters between healthy controls and individuals with dizziness.
Full size table

Finally, step length was slightly decreased in the dizziness group compared to controls, but the difference was not significant (0.71 ± 0.08 m vs. 0.63 ± 0.06 m, p = 0.153). However, the residuals did not pass the homoscedasticity and independence tests (see Figure S3 in supplementary material).

Discussion

Falls are a leading cause of mobility loss, hospitalization, and reduced quality of life, particularly in older adults. Quantitative gait assessment is a critical tool for identifying individuals at heightened risk of falls20. For example, decreased gait speed, prolonged stance phase and double support duration, and reduced cadence are well-established markers of increased fall risk21,22,23,24,25. However, obtaining such gait measures typically requires costly sensors, controlled environments, and highly trained personnel. As a result, quantitative gait assessments remain largely confined to research settings or specialized rehabilitation clinics, severely limiting their accessibility for routine clinical care and large-scale screening in community or primary care settings.

VisionMD-Gait addresses these critical barriers by providing a scalable, accessible, and user-friendly platform for clinical-grade spatiotemporal gait analysis using only frontal-view videos recorded with a smartphone in standard clinical settings. Our results found high correlation and agreement between video-derived measures and wearable sensor data across multiple populations, including healthy controls and persons with dizziness, and gait parameters, including gait speed, cadence, step duration, stride duration, stance time, swing time, double support time, and step length. Moreover, mean absolute errors were lower than 10% of the sensor-bases estimates for all measures. These results confirm that VisionMD-Gait provides accurate and clinically relevant gait metrics that agree with traditional sensor-based approaches.

Furthermore, our clinical analysis demonstrated that VisionMD-Gait can identify significant gait alterations in individuals with dizziness when compared to healthy controls, including slower gait speed, reduced cadence, increased step duration, stride duration, and prolonged stance and double support time. These findings are consistent with a cautious gait strategy, characterized by deliberate, slower limb movements. Such adaptations have been documented in vestibular-compromised individuals and are recognized predictors of reduced mobility and elevated fall risk3,22,23,24,26. Importantly, this demonstrates VisionMD-Gait’s potential to uncover subclinical gait impairments in individuals who do not report overt gait complaints but may still be at elevated risk for falls. VisionMD-Gait’s capabilities could be transformative for prevention efforts and for evaluating the effectiveness of interventions aimed at improving gait in both clinical and community settings.

Our results align with a growing body of literature validating monocular, frontal-view video approaches for gait analysis. Azhand et al. (2021) reported high agreement between algorithmically extracted gait measures from frontal-view videos and an instrumented walkway27. Barzyk et al. (2024) validated a smartphone camera approach in stroke patients, achieving spatiotemporal metrics comparable to those provided by a Motion Capture System28. Stenum et al. (2024) demonstrated that video-based systems can capture clinically relevant, condition-specific gait parameters and track within-participant changes over time14. Similarly, Cimorelli et al. (2024) introduced a portable, in-clinic system for video-based quantitative gait assessments, demonstrating high accuracy in prosthetic users compared with wearable sensors15.

VisionMD-Gait builds on these advances but introduces key innovations. Unlike most existing methods that require technical expertise for setup and data analysis, VisionMD-Gait provides an end-to-end solution that does not require any programming or specialized knowledge and processes all data locally without using any external servers or cloud processing solutions. This eliminates the need to share sensitive patient data (gait videos) with third parties, addressing critical privacy and data security concerns. Moreover, VisionMD-Gait is platform-agnostic and can run on standard computers found in research or clinical settings, making it highly scalable and adaptable for use in diverse applications, allowing clinicians, nurses, technicians, and even patients themselves to collect and analyze gait data without specialized training or infrastructure.

Despite its promise, VisionMD-Gait has some limitations. First, while the algorithm was accurate across participants, slight inaccuracies may occur in cases of poor lighting or occlusions. Future efforts will focus on refining the algorithm for challenging recording conditions and expanding validation to include diverse clinical populations. Additionally, our current analysis excluded turning events to focus on straight walking segments. Incorporating turning metrics might provide additional clinical information and is an important next step to further enhance VisionMD-Gait clinical utility29.

In summary, VisionMD-Gait represents a transformative step toward democratizing access to quantitative gait assessments. This tool enables clinicians and researchers to obtain accurate, objective gait parameters without the need for expensive hardware or technical expertise. VisionMD-Gait has the potential to facilitate identification and quantification of mobility impairments, personalize rehabilitation strategies, and assess the effect of gait therapies from smartphone videos, which can be easily captured. Its ability to detect gait differences in patients with dizziness when compared to healthy controls highlights its promise for screening and monitoring. The integration of VisionMD-Gait into routine practice could profoundly impact the detection and management of gait abnormalities, fall prevention strategies, and research on mobility disorders and aging.

Methods

Participants

This study recruited 24 healthy participants and 10 patients with vestibular dizziness. All participants were recruited through the vestibular clinic at UF Health ENT and Allergy – The Oaks. Participants in the dizziness group were either diagnosed with benign paroxysmal positional vertigo (BPPV) or reported symptoms consistent with vestibular dizziness. Healthy controls were recruited from relatives of patients and from other individuals seen at the same clinic without vestibular complaints. Participants in the dizziness cohort were excluded if their dizziness or imbalance was attributable to orthostatic hypotension, stroke, Parkinson’s disease, cerebral atrophy, paralysis, or other neurological or neuromuscular disorders. Healthy control participants were included only if they had no self-reported history of orthopedic or neurological disorders or any other disorders that affects gait and did not require the use of a cane, walker, or armchair for support.

This study was conducted in accordance with the principles of the Declaration of Helsinki, and all participants provided written informed consent prior to enrollment.

Data acquisition

Gait events were recorded using a smartphone (iPhone 12) at 60 frames per second (fps) with a resolution of 1080 × 1920 pixels. To validate the markerless method, a wearable 3D motion capture system, Noraxon Ultium MotionTM IMUs (Noraxon, Inc., Scottsdale, AZ), was used simultaneously to collect kinematic gait data. This system employs multiple inertial measurement units (IMUs) placed on various body segments to track acceleration, movement, and three-dimensional (3D) position, enabling detailed analysis of gait kinematics.

Figure 2A) illustrates the wearable sensors arrangement; participants wore sensors placed bilaterally on their feet, shanks, thighs, and a single sensor on the lower back to capture comprehensive gait metrics. Figure 2B) presents a schematic of the data collection setup, depicting an 8-meter walkway with designated start and end points and the position of the smartphone camera used for video data acquisition. Participants were instructed to walk at a comfortable, self-selected speed, turn upon reaching the end of the walkway, and complete at least four laps, depending on their health status and comfort level, but were encouraged to complete more. Videos and wearable sensor data were acquired concurrently and were subsequently segmented to ensure alignment between modalities. Video segmentation was performed directly within VisionMD-Gait, excluding turns and retaining only steady-state walking intervals. Sensor data were segmented using the manufacturer’s software (MR3.20, Noraxon Inc., Scottsdale, AZ), and segments were selected to match the corresponding video intervals, ensuring that both datasets captured the same gait events and contained the same number of steps.

Fig. 2
figure 2

Validation study setup and clinical environment. (A) Placement of wearable inertial sensors on the lower back, thighs, shanks, and feet for estimation of ground truth gait measurements. (B) Schematic of the clinical hallway used for video and sensor data acquisition. Participants walked an 8-meter path while a smartphone camera recorded from a fixed position. (C) Real-world recording setup showing smartphone camera placement and the participant performing the gait task.

Full size image

Figure 2C) shows the real-world recording setup employed in this study. All data were acquired in a vestibular clinic hallway (UF Health The Oaks, Gainesville, FL) using a smartphone supported on a tripod, positioned behind the participant at a distance sufficient to capture their full body. The video acquisition workflow required no adjustments to the clinic environment or patient routines.

Data analysis

All gait videos were analyzed using the VisionMD-Gait processing pipeline; Fig. 3 illustrates the full processing workflow of VisionMD-Gait which consists of three main components: a pose estimation model, a gait transformer model, and a feature extractor. Pose estimation was performed using MeTRAbs, a monocular 3D human pose estimator model trained on 28 open-sourced datasets with high-quality 3D human pose labels19. MeTRAbs receives as input the video frames along with a bounding box indicating the position of the subject in the frame, which was estimated using a YOLO-based object detector trained to identify human bodies. The model’s output is the 3D position of multiple joint coordinates in metrics units with the position of the camera as the origin.

Fig. 3
figure 3

Overview of the VisionMD-Gait processing pipeline. Original video frames are processed to reconstruct 3D human poses, which are then passed through a gait transformer model to extract joint kinematics and gait events. From these outputs, spatiotemporal gait parameters such as gait speed, cadence, step length, stance time, and swing time are estimated.

Full size image

The resulting joint coordinates were re-ordered to align with the Human3.6 M Dataset format and were re-centered at the pelvis30. These standardized joint sequences, along with the participant’s height, were passed to the gait transformer model, which estimated lower-limb joint kinematics (hip, knee, and ankle angular displacement and velocities) and detected foot contact events, including heel strikes and toe-offs15. Gait phase parameters, including stance time, swing time, and double support time, were computed based on the timing of these gait events. Cadence was computed based on the number of steps estimated in the video. Other parameters, including and gait speed were derived from changes in the estimated distance of the pelvis to the camera at each video frame provided by MeTRAbs. Table 4 summarizes the spatial, temporal, and spatiotemporal gait parameters derived from videos processed with VisionMD-Gait, including gait speed, cadence, step duration, stride duration, stance time, swing time, double support time, and step length.

Table 4 Description of gait parameter computations. Gait features were computed using heel strike, toe-off timestamps, and depth data from pose estimation. Metrics follow standard definitions to enable consistent comparison across methods.
Full size table

To reduce potential viewpoint-related asymmetries introduced by pose-estimation biases (rather than correcting true physiological asymmetry), each video was vertically mirrored and reprocessed using the full analysis pipeline. Because mirroring swaps the visual left and right sides, gait events were reassigned accordingly to preserve each limb’s true identity. Final gait parameter values for each limb were obtained by averaging the corresponding outputs from the original and mirrored videos. Importantly, this procedure does not eliminate or alter left–right asymmetries in the gait pattern; instead, it helps ensure that any side differences detected arise from the participant’s movement rather than from model biases.

Reference kinematic gait parameters were extracted from the wearable sensors using the MR3.20 software (Noraxon, Inc., Scottsdale, AZ). To assess the validity of video-based gait estimates, we first computed Pearson’s correlation coefficients between sensor-derived and video-derived measures, we then quantified accuracy using the mean absolute error (MAE%), expressed relative to the sensor-based estimates. Finally, agreement between sensor-derived and video-derived measures was evaluated using Bland-Altman analysis. All analyses were performed separately for the healthy control and vestibular dizziness groups to demonstrate that the pipeline provides accurate and reliable results across distinct populations.

To compare kinematic gait differences between participants with vestibular dizziness and healthy controls, we employed a Linear Mixed Model (LMM), which appropriately handles repeated measures and accounts for both fixed and random effects. Participant age, gender, and height were included as covariates to account for demographic variability. Because each participant completed multiple walking trials, we incorporated random intercepts to account for individual baseline differences. Equation 1 summarizes the model structure, where j indexes participants and i indexes repeated trials within participant j. y​ij denotes the gait measure (e.g., step time, step length) for the i-th trial of the j-th participant. Groupij​ is a binary variable (0 = healthy control, 1 = vestibular dizziness). u₀j is the random intercept for participant j, and εij is the residual error:

$$\:{y}_{ij}={\beta\:}_{0}+{\beta\:}_{1}{Group}_{ij}+{\beta\:}_{2}Ag{e}_{ij}+{\beta\:}_{3}{Gender}_{ij}+{\beta\:}_{4}{Height}_{ij}+{u}_{0j}+{ϵ}_{ij}$$
(1)

Each β coefficient was tested against the null hypothesis of no effect, with statistical significance defined as p < 0.05. All p-values were corrected for multiple comparisons using the Bonferroni method. Model assumptions were assessed by examining residuals to ensure normality (Shapiro–Wilk test), homoscedasticity (Breusch–Pagan test), and independence (Ljung–Box test). All analyses were performed in Python using the statsmodels package, models were fitted via restricted maximum likelihood estimation using the L-BFGS optimizer.