Document Type : Research Paper
Authors
1
Department of Physical Education and Sports Science, Faculty of Psychology and Educational Sciences, Yazd University, Yazd, Iran
2
Department of Physical Education, Farhangian University, P.O. Box 889-14665, Tehran, Iran
3
. Department of Sport Injuries and Biomechanics, Faculty of Sports and Health Sciences, Tehran University, Tehran, Iran
Abstract
Background and Purpose
Motion analysis is essential for understanding human movement, with traditional approaches relying heavily on marker-based optical systems valued for their high accuracy and temporal resolution. Nonetheless, these systems are typically costly, time-consuming, and confined to controlled laboratory environments. Recent advancements have shifted attention towards more practical alternatives, including reduced-marker and markerless motion capture technologies. Reducing marker count streamlines setup while still capturing critical gait parameters. Markerless systems, empowered by deep learning and computer vision, enable natural, unrestricted gait assessments in real-world contexts. Despite some limitations in distal segment accuracy, these approaches are increasingly viable in research and clinical applications. Deep neural networks are central to this progress by facilitating automated anatomical landmark tracking. This study aims to validate a neural network model that estimates step length during walking using only three anatomical markers positioned on the pelvis, trunk, and head, balancing data simplicity and predictive efficacy.
Methods
Twenty-eight healthy young males (22.2 ± 2.09 years), recruited via purposive sampling, participated after providing informed consent; the protocol was approved by Yazd University Institutional Review Board (IR.YAZD.REC.1403.078). Participants performed treadmill walking at three speeds (preferred, slow, fast) for minimum durations of 40 seconds per speed while being recorded by a 3D motion capture system (Optitrack Duo120, USA) sampling at 120 Hz. Reflective marker clusters were placed on participants’ pelvis and back of the head; single markers were positioned on the C7 vertebra and heels. Marker positions were averaged to extract linear displacement data for head, trunk, and pelvis.
Raw data were filtered with a fourth-order Butterworth low-pass filter (6 Hz cutoff) and analyzed in MATLAB (R2021b). A custom script converted local to global coordinates, segmented gait cycles, normalized marker positions relative to participant height, and extracted step length at initial double support phase. Final data matrices comprised 4,710 time-normalized frames per marker. Thirteen predictor sets combining 3 to 9 features across sagittal, frontal, and transverse planes from pelvis, trunk, and head markers were created.
A two-layer feedforward neural network with 20 hidden neurons was trained using Levenberg–Marquardt optimization. Data were partitioned into training (70%), validation (15%), and test (15%) subsets. Step length served as the target variable. Model performance was appraised by Bland–Altman plots, mean error, coefficient of variation (CV), root mean square error (RMSE), and Pearson correlation coefficients — interpreted as weak (≤0.40), moderate (0.40–0.74), or strong (≥0.75). Analyses were conducted using MATLAB custom scripts.
Results
Mean step length based on heel markers was 29.8% of participant height (SD = 1.5%). From 4,710 gait cycles, 707 (15%) random samples validated the model predictions, with statistical characteristics mirroring the full dataset. The top-performing model utilized six features representing frontal and sagittal plane movements of all three markers, achieving a strong correlation with direct step length measures (R = 0.93), coefficient of determination (R²) of 86.5%, and RMSE of 7.59 cm. The poorest performance occurred when using only lateral marker movements (R = 0.31). Ten out of thirteen predictor configurations demonstrated strong to very strong correlations, all statistically significant, reinforcing the feasibility of accurate step length estimation with minimal upper-body markers.
Conclusion
This study validates a neural network-based approach to estimate step length during walking using just three anatomical markers on the pelvis, trunk, and head. The highest accuracy arose when incorporating anterior-posterior and superior-inferior directional data, aligning with biomechanical principles where forward progression and vertical center-of-mass oscillations predominate gait dynamics. Models incorporating medial-lateral data alone exhibited poorer performance due to minimal variance in those directions. Integrating multiple markers and movement planes yields comprehensive gait characterization, while reliance on single markers suffers from reduced informational content and increased signal variability, underscoring the benefit of multi-input neural network models.
These results concur with prior studies employing machine learning or wearable sensor methods (IMUs), which reported correlations ranging from 0.79 to 0.95 for kinematic and kinetic estimates. Marker-based approaches without wearables similarly showed strong correlations (0.85–0.94), dependent on marker placement and target variables.
The utilization of central body anatomical landmarks, more reliably detected in markerless systems compared to distal limb markers, represents a meaningful advancement in simplifying gait analysis, reducing equipment needs, and accelerating data processing. This method holds promise especially in clinical and sports settings constrained by resource limitations.
Article Message
Step length estimation via a neural network trained on minimal motion data from pelvis, trunk, and head markers in primary movement directions offers high predictive validity relative to traditional heel-marker measures. Leveraging central anatomical landmarks enhances practicality and cost-effectiveness of gait assessment without sacrificing accuracy, facilitating widespread adoption in diverse applied settings.
Ethical Considerations
The study was approved by the Institutional Review Board of Yazd University (Approval Code: IR.YAZD.REC.1403.078). Participant consent and ethical guidelines were rigorously observed.
Authors’ Contributions
Conceptualization: Mostafa Haj Lotfalian (33%), Mohammad Javad Razi (33%), Fatemeh Zare Bidoki (33%)
Data Collection: Mostafa Haj Lotfalian (60%), Mohammad Javad Razi (20%), Fatemeh Zare Bidoki (20%)
Data Analysis: Mostafa Haj Lotfalian (40%), Mohammad Javad Razi (30%), Fatemeh Zare Bidoki (30%)
Manuscript Writing: Mostafa Haj Lotfalian (33%), Mohammad Javad Razi (33%), Fatemeh Zare Bidoki (33%)
Review and Editing: Mostafa Haj Lotfalian (30%), Mohammad Javad Razi (35%), Fatemeh Zare Bidoki (35%)
Literature Review: Mohammad Javad Razi (50%), Fatemeh Zare Bidoki (50%)
Project Management: Mostafa Haj Lotfalian (100%)
Conflict of Interest
The authors declare no conflicts of interest related to this article.
Acknowledgments
The authors sincerely thank all participants for their valuable contributions to this study.
Keywords
Subjects
