Journal of Biomechanics 155 (2023) 111645

Contents lists available at ScienceDirect

Journal of Biomechanics
journal homepage: www.elsevier.com/locate/jbiomech

Validity of artificial intelligence-based markerless motion capture system

for clinical gait analysis: Spatiotemporal results in healthy adults and adults
with Parkinson’s disease
Zachary Ripic a, c, Joseph F. Signorile a, b, Thomas M. Best c, d, Kevin A. Jacobs a, Mitch Nienhuis a,
Cole Whitelaw a, Caden Moenning a, Moataz Eltoukhy a, e, *
a
Department of Kinesiology and Sport Sciences, University of Miami, Miami, FL, United States
b
Center on Aging, University of Miami Miller School of Medicine, Miami, FL, United States
c
Sports Medicine Institute, University of Miami Miller School of Medicine, Miami, FL, United States
d
Department of Orthopaedics, University of Miami Miller School of Medicine, Miami, FL, United States
e
Department of Industrial and Systems Engineering, University of Miami, Miami, FL, United States

A R T I C L E I N F O A B S T R A C T

Keywords: Markerless motion capture methods are continuously in development to target limitations encountered in
Gait analysis marker-, sensor-, or depth-based systems. Previous evaluation of the KinaTrax markerless system was limited by
Human pose estimation differences in model definitions, gait event methods, and a homogenous subject sample. The purpose of this work
Markerless motion capture
was to evaluate the accuracy of spatiotemporal parameters in the markerless system with an updated markerless
Artificial intelligence
Computer vision
model, coordinate- and velocity-based gait events, and subjects representing young adult, older adult, and
Parkinson’s disease groups. Fifty-seven subjects and 216 trials were included in this analysis. Interclass corre­
lation coefficients showed excellent agreement between the markerless system and a marker-based reference
system for all spatial parameters. Temporal variables were similar, except swing time which showed good
agreement. Concordance correlation coefficients were similar with all but swing time showing moderate to
almost perfect concordance. Bland-Altman bias and limits of agreement (LOA) were small and improved from
previous evaluations. Parameters showed similar agreement across coordinate- and velocity-based gait methods
with the latter showing generally smaller LOAs. Improvements in spatiotemporal parameters in the present
evaluation was due to inclusion of keypoints at the calcanei in the markerless model. Consistency in the calcanei
keypoints relative to heel marker placements may improve results further. Similar to previous work, LOAs are
within boundaries to detect differences in clinical groups. Results support the use of the markerless system for
estimation of spatiotemporal parameters across age and clinical groups, but caution should be taken when
generalizing findings due to remaining error in kinematic gait event methods.

1. Introduction minimally restrictive, allow for measurements in a natural environ­

ment, and have a sufficiently large field of view (Mündermann et al.,
Understanding how underlying systems of the human body interact 2006). Optical marker-based systems typically have millimeter level
to produce motion in normal or pathological states requires appropriate accuracy (Topley and Richards, 2020) but several limitations reduce
measurement, description, and analysis (Lu and Chang, 2012; Winter, their validity and ease of implementation in practice. Alternative sys­
2009). Often direct measurements are difficult, or not possible, and tems using inertial measurement units (IMUs) (Zago et al., 2018) or
musculoskeletal modeling could help understand the interaction be­ integrated RGB images and depth sensors (RGB-D) (Eltoukhy et al.,
tween the different elements and the resulting movement biomechanics 2017; Oh et al., 2020) have been proposed, both approaches however
both in healthy and pathological conditions (Donno et al., 2022). have their own set of limitations including depth of view, integration
However, the development of such models requires some in vivo mea­ drift, sampling rate, and accuracy concerns (Guffanti et al., 2020;
surements. An ideal measurement system should be non-invasive, Kumarasiri et al., 2018; Razavian et al., 2019; Yeung et al., 2021).

* Corresponding author at: Department of Kinesiology and Sport Sciences, University of Miami, Miami, FL, United States.
E-mail address: meltoukhy@miami.edu (M. Eltoukhy).

https://doi.org/10.1016/j.jbiomech.2023.111645
Accepted 16 May 2023
Available online 19 May 2023
0021-9290/© 2023 Elsevier Ltd. All rights reserved.
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

Therefore, there is a need for innovative systems which encompass without assistance, 2) body mass index > 35 kg•m− 2, 3) adults unable
methods that overcome the above limitations. to consent (<23 on the MoCA), and 4) PD subjects who present with
Markerless motion capture methods using concepts in deep learning, severe disability or are wheelchair bound (H&Y stages 4, or 5).
particularly convolutional neural networks (CNNs) (Goodfellow et al.,
2016), have outperformed classic computer vision algorithms on various 2.2. Procedure
tasks (Krizhevsky et al., 2012; Ren et al., 2015; Shelhamer et al., 2017).
Importantly, CNNs have become an essential component for body part This study was approved by the University’s Institutional Review
estimation in two-dimensional (2D) and three-dimensional (3D) human Board. Using a cross-sectional design, subjects were brought into the
pose estimation (HPE) algorithms from RGB inputs (He et al., 2016; laboratory one time for approximately 30 min. The testing protocol and
Huang et al., 2017; Newell et al., 2016; Sandler et al., 2018; Tan and Le, the benefits and risks of the study were explained, and informed consent
2019). Various monocular and multi-view approaches exist but the was obtained prior to testing. Following informed consent, in those over
latter tend to perform better when evaluated on benchmark datasets 50 years of age, subjects were asked to complete the MoCA test. Subjects
(Desmarais et al., 2021). were screened on their recent physical activity and asked to report any
Few studies have been conducted on the validity of available multi- unusual soreness or pain that may affect results on their day of visit.
view markerless systems outside of benchmark datasets, and even fewer Height and mass were measured before testing. Subjects were asked to
have targeted 3D kinematics during gait. Several studies have compared change into provided spandex clothing if their clothing upon arrival was
gait spatiotemporal variables between marker-based and markerless likely to cause error in the marker-based measurement. Twenty-six
approaches (Kanko et al., 2021; Moro et al., 2022) or markerless reflective markers were placed according to a modified Plug-in Gait
compared to GaitRite (Kanko et al., 2021; Lonini et al., 2022) and marker set (Vicon Motion Systems Inc., Oxford, UK). Additional markers
sensor-based (Mehdizadeh et al., 2021) methods. Previous results for the were placed bilaterally on the medial femoral condyles, medial malleoli,
KinaTrax (KinaTrax Inc., Boca Raton, FL) markerless system showed antero-inferior thighs, antero-inferior shanks, and on the proximal fifth
excellent agreement between most spatial parameters calculated using metatarsophalangeal joints to aid marker processing steps and 3D
coordinate-based gait events (Zeni et al., 2008) compared to those segment reconstruction.
calculated from marker-based tracking and force plate events (Ripic Gait trials were performed along a 10-m walkway in the laboratory
et al., 2022). Worse agreement was found in temporal parameters. This equipped with an eight-camera marker-based system (SMART-DX; BTS
prior error between approaches was influenced by inconsistencies be­ Bioengineering, Milan, Italy) and eight-camera markerless system
tween models and potentially due to coordinate-based gait events (Ripic (KinaTrax Inc., Boca Raton, FL). Both systems recorded at 100 frames
et al., 2022) without the explicit comparison of other gait event per second (fps). Prior to each subject visit the marker-based system was
methods. However, these two limitations are not independent from each calibrated. Calibrations were adequate when average re-projection error
other, and further work was warranted to explore the effect of gait event for all cameras was less than 0.5-mm. To initialize the markerless system
methods and their input variables. Importantly, only the ankle joint a single background image of the capture volume was obtained before
center was available in the previous markerless model to determine static trials and used in post-processing steps.
heel-strike whereas the heel marker was used in the marker-based Before dynamic trials a single static trial was obtained for the
model. Since the gait events are determined from peak positions of the marker-based system with the subject in the center of the capture vol­
proximal and distal foot segments relative to the pelvis in the ante­ ume. Reconstructed static trials were visually inspected to ensure all
roposterior direction, differences in input values from the two models experimental markers were present and that the volume was absent of
could largely influence the event timing and subsequent parameters any additional reflective surfaces. Prior to dynamic trials, subjects
derived from these events. Additionally, the original work was done in received instructions to walk at their preferred pace along the walkway.
only younger individuals, therefore performance of the markerless sys­ Practice trials were given to ensure that one gait cycle per side was
tem in older adults and other groups remained unexplored. recorded within the experimental area considered in post-processing.
The purpose of this study was to evaluate spatiotemporal parameters Following, dynamic trials were conducted with subjects starting at
calculated with the latest version of the KinaTrax markerless model, their determined starting position and walking in a straight line at their
which was retrained to include keypoints on the calcanei to improve preferred pace along the walkway. If one full gait cycle in each trial was
consistency with the marker-based model. The project also aimed to not achieved, trials were repeated until three successful trials were ob­
address prior issues in gait event and sampling methods. We hypothe­ tained for up to 10 total trials. Adequate rest was given between trials
sized that spatiotemporal parameters would show excellent agreement and upon subject request.
and consistency, nearly perfect concordance, and small bias and limits of
agreement compared to a marker-based system. Additionally, better 2.3. Data analysis
agreement was expected when comparing parameters calculated from
velocity-based gait events. Marker-based motion files were transferred from the collection sys­
tem to an offline processing station. All marker trajectories were pro­
2. Methods cessed using Nexus v.2.12.0 (Vicon Motion Systems Inc., Oxford, UK)
with a previously built labeling skeleton. Following marker processing,
2.1. Participants trials were input to Visual 3D (C-Motion Inc., Germantown, MD) for
skeletal modeling, gait event determination, and spatiotemporal
A convenience sample of 57 (Young Adult: N = 23, Older Adult: N = parameter estimation.
14, PD: N = 20) subjects between the ages of 18–85 were included in this Markerless processing procedures have been outlined previously
study. Subjects were recruited from existing databases, ongoing inter­ (Ripic et al., 2022). Data collection sessions were adjusted for any
vention studies, the University community, and surrounding area. camera drift by aligning the static background image in each camera
Subjects were included if they met the following criteria: 1) able to view from the current session with the latest system calibration back­
ambulate 10-m without assistance, 2) no prior spinal cord injury, and 3) ground images. This quality assurance step corrects any minor changes
no musculoskeletal injury in the previous six months. Adults with Par­ in camera orientation that might occur between calibration days, as the
kinson’s disease (PD) were included if they met the following additional calibration is not done locally prior to each session. After drift correc­
criteria: 1) mild to moderate impairment (H&Y stages 1–3) and 2) score tion, videos from the current session were compiled and processed using
23 or above on the Montreal Cognitive Assessment (MoCA) (Thomann the HumanVersion3 model from KinaTrax. HumanVersion3 is a full-body
et al., 2020). Exclusion criteria included: 1) unable to ambulate 10-m model containing 40 landmarks with newly trained points added

2
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

bilaterally on the lower body containing the calcaneus, first metatarsal 0.5 indicate poor validity, values of 0.5–0.75 indicate moderate validity,
head, fifth metatarsal head, hallux head, and the patella center. The values of 0.75–0.9 indicate good validity, and values greater than 0.9
model was re-trained on over 500 frames (>4000 images) from videos indicate excellent validity (Koo and Li, 2016). Additionally, Bland-
collected in the laboratory. Videos and frames were selected to include a Altman bias, confidence intervals, and limits of agreement were calcu­
variety of poses and were manually annotated with additional key­ lated for all spatiotemporal parameters to determine if the proposed
points. After being processed locally on the KinaTrax system, segment markerless approach yielded errors within acceptable limits and would
poses from the markerless system were input to Visual 3D for skeletal be appropriate for clinical use.
modeling, gait event determination, and spatiotemporal parameter
estimation. 3. Results
Previously, kinematically-determined gait events indicated that the
method using position of the foot segments relative to the sacrum (Zeni A combined 215 trials from young adult (M/F: 16/7, Age: 21.83 ±
et al., 2008) resulted in good estimations of spatial gait parameters but 2.37 years, Height: 1.76 ± 0.09 m, Mass: 77.07 ± 10.71 kg, BMI: 24.73
was likely a source of error when estimating swing time and double limb ± 2.70 kg•m− 2), older adult (M/F: 6/8, Age: 75.21 ± 5.49 years, Height:
support (DLS) duration. Alternatively, Zeni et al. (2008), proposed a 1.66 ± 0.15 m, Mass: 79.56 ± 16.00 kg, BMI: 29.15 ± 6.18 kg•m− 2,
second method to determine gait events using the velocity of the prox­ MoCA: 27.64 ± 1.65), and PD (M/F: 15/5, Age: 68.70 ± 8.37 years,
imal and distal foot segments during treadmill walking where the zero- Height: 1.69 ± 0.07 m, Mass: 74.57 ± 14.14 kg, BMI: 26.11 ± 4.25
crossing of the anteroposterior velocity vector indicates the HS and TO kg•m− 2, MoCA: 27.25 ± 2.17, Hoehn & Yahr Stage: 2.0 ± 1.0) groups
events (Zeni et al., 2008). The velocity-based algorithm showed better were included in the analysis. ICC between systems and methods are
results in both healthy and pathological groups but both methods were shown in Tables 2 and 3. Spatiotemporal parameters (means ± SD) for
within ±2 frames of force plate measured gait events at 60 fps. The the marker-based and markerless systems calculated using both velocity-
velocity-based coordinate method was generally < ± 1 frame error in and coordinate-based gait events are shown in Tables 4 and 5 along with
healthy and clinical groups (Zeni et al., 2008). A recent study comparing the results of the Bland-Altman analysis (example given in Fig. 1). Re­
the accuracy of several kinematic methods to gold standard force plate sults by group are provided in the Appendix.
measurements indicated that the velocity-based approach was the best For spatiotemporal parameters calculated using velocity-based gait
performing method (Zahradka et al., 2020), but given the opportunity to events, ICCs for absolute agreement were excellent in all spatial pa­
expand on previous results, both methods were applied within Visual 3D rameters. Nearly identical results were found when spatial parameters
for marker-based and markerless tracking. were calculated using coordinate-based gait events. For temporal gait
variables, all showed excellent agreement when calculated using
2.4. Statistical analysis velocity-based gait events. Similar results were found when parameters
were calculated using coordinate-based gait events, except for swing
For the spatiotemporal validation, a total of 24 variables (including time which was only slightly lower compared to velocity-based
left and right) were extracted from and compared across all trials be­ calculations.
tween the marker-based and markerless systems. Descriptions for the 10
base parameters are given in Table 1. Linearity between variables
calculated from each system was checked using scatterplots. Table 2
The absolute agreement and relative consistency between systems Two-Way Mixed Effects, single measurement, Interclass Correlation Coefficients
were assessed with interclass correlation coefficients (ICCs). ICC esti­ for absolute agreement for all spatiotemporal parameters obtained from marker-
mates and their 95% confidence intervals (CI95%) were calculated using based and markerless tracking using both velocity- and coordinate-based gait
events for calculations.
SPSS statistical package version 28 (SPSS Inc., Chicago, IL) based on a
single measurement, two-way mixed effects model. ICC values less than Velocity Coordinate

Parameter ICC Low High ICC Low High

Table 1 Speed (m/s) 0.995 0.993 0.996 0.995 0.994 0.996
Description of spatiotemporal parameters calculated from gait events in Visual Stride Length (m) 0.992 0.990 0.994 0.991 0.988 0.993
3D for both marker-based and markerless systems. Stride Width (m) 0.980 0.974 0.985 0.983 0.978 0.988
Step Length (m) 0.984 0.979 0.988 0.981 0.975 0.985
Gait Parameter Description
Cycle Time (s) 0.990 0.987 0.992 0.987 0.983 0.990
Swing Time (s) Period while the foot is not in contact with the ground (TO-HS Stance Time (s) 0.972 0.963 0.979 0.966 0.956 0.974
in same leg). Swing Time (s) 0.910 0.884 0.931 0.888 0.856 0.913
Stance Time (s) Period when the foot is in contact with the ground (HS-TO in Step Time (s) 0.978 0.971 0.984 0.974 0.966 0.980
same leg). DLS (s) 0.931 0.911 0.947 0.917 0.893 0.936
Double Support Initial and final part of stance phase when two feet are in Cadence (strides/min) 0.985 0.980 0.988 0.983 0.977 0.987
Time (s) contact with the ground (global time = initial + final double Left
support). Stance Time (s) 0.965 0.954 0.973 0.956 0.942 0.966
Step Time (s) Period taken for one step measured from HS of one foot to HS Step Length (m) 0.981 0.975 0.986 0.978 0.971 0.983
of the other foot. Step Time (s) 0.977 0.970 0.983 0.973 0.965 0.979
Stride/Cycle Time Total amount of time for the cycle measured between Step Cadence (steps/min) 0.976 0.969 0.982 0.973 0.965 0.979
(s) successive HS in the same leg. Stride Length (m) 0.992 0.989 0.994 0.988 0.984 0.991
Cadence (strides/ Number of strides taken in a minute. Stride Cadence (Strides/ 0.980 0.974 0.985 0.977 0.970 0.982
min) min)
Stride Length (m) Distance travelled during one stride (or cycle); Distance Swing Time (s) 0.899 0.870 0.922 0.877 0.842 0.905
between two successive HS in the same leg Right
Step Length (m) Distance between corresponding HS in both legs; HS to HS in Stance Time (s) 0.979 0.972 0.984 0.976 0.969 0.982
opposite leg Step Length (m) 0.987 0.983 0.990 0.983 0.978 0.987
Speed (m/s) Distance covered per second; Calculated as the measured Step Time (s) 0.979 0.972 0.984 0.975 0.967 0.981
stride length divided by the measured stride time Step Cadence (steps/min) 0.978 0.971 0.983 0.978 0.971 0.984
Stride Width (cm) Perpendicular distance (frontal plane) between the position of Stride Length (m) 0.984 0.979 0.988 0.982 0.976 0.986
the proximal ipsilateral foot at ipsilateral HS to the position of Stride Cadence (Strides/ 0.989 0.985 0.991 0.988 0.984 0.991
the proximal end position of the foot at the next contralateral min)
HS Swing Time (s) 0.920 0.897 0.939 0.898 0.869 0.921

HS: Heel-strike, TO: Toe-off. ICC: Interclass correlation coefficients, DLS: Double Limb Support Time.

3
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

Table 3 interval for bias neared the line of equality for stride width indicating
Two-Way Mixed Effects, single measurement, Interclass Correlation Coefficients potential overestimation in the markerless system. Similarly, the line of
for relative consistency for all spatiotemporal parameters obtained from marker- equality was not in the interval for left stride length, indicating that this
based and markerless tracking using both velocity- and coordinate-based gait measure was underestimated by the markerless system. Similar results
events for calculations. were found when comparing the marker-based and markerless spatio­
Velocity Coordinate temporal parameters using the coordinate-based method, except no
Parameter ICC Low High ICC Low High underestimation was noted for left stride length. Additionally, the line of
equality was not in the confidence interval for right step cadence,
Speed (m/s) 0.995 0.993 0.996 0.995 0.994 0.996
Stride Length (m) 0.992 0.990 0.994 0.991 0.988 0.993 indicating a small overestimation in the markerless system.
Stride Width (m) 0.981 0.975 0.985 0.984 0.979 0.988 When expressed as a percentage of the spatiotemporal parameter
Step Length (m) 0.984 0.979 0.988 0.981 0.975 0.985 means from the marker-based system, bias was less than one percent
Cycle Time (s) 0.990 0.987 0.992 0.987 0.984 0.990 with less than eight percent deviation in the markerless system for both
Stance Time (s) 0.947 0.932 0.960 0.967 0.956 0.974
Swing Time (s) 0.910 0.883 0.930 0.888 0.856 0.913
velocity-based and coordinate-based methods.
Step Time (s) 0.978 0.972 0.983 0.975 0.967 0.980
DLS (s) 0.860 0.821 0.891 0.917 0.892 0.936 4. Discussion
Cadence (strides/min) 0.985 0.980 0.988 0.983 0.977 0.987
Left
The purpose of the spatiotemporal validation was to evaluate the
Stance Time (s) 0.964 0.954 0.973 0.956 0.942 0.966
Step Length (m) 0.981 0.975 0.985 0.978 0.971 0.983 validity of spatiotemporal variables calculated from markerless tracking
Step Time (s) 0.977 0.970 0.982 0.973 0.965 0.979 with an updated markerless model, methodological comparisons for gait
Step Cadence (steps/min) 0.976 0.968 0.981 0.973 0.965 0.979 event detection, and a more heterogenous sample across age and PD.
Stride Length (m) 0.992 0.990 0.994 0.989 0.985 0.991 The hypothesis that the markerless system would show excellent
Stride Cadence (strides/ 0.980 0.974 0.985 0.977 0.970 0.983
min)
agreement and small bias with narrow LOA is supported by most results.
Swing Time (s) 0.899 0.869 0.922 0.877 0.842 0.905 ICC for absolute agreement showed excellent agreement between the
Right marker-based and markerless systems for both velocity- and coordinate-
Stance Time (s) 0.930 0.910 0.946 0.977 0.970 0.982 based spatiotemporal calculations. The only parameters not reaching
Step Length (m) 0.987 0.983 0.990 0.984 0.979 0.987
excellent agreement were swing time (coordinate – absolute agreement)
Step Time (s) 0.979 0.973 0.984 0.976 0.968 0.981
Step Cadence (steps/min) 0.978 0.971 0.983 0.979 0.973 0.984 and double limb support time (velocity – relative consistency), which
Stride Length (m) 0.984 0.979 0.988 0.982 0.976 0.986 showed good agreement and were consistent between event methods.
Stride Cadence (strides/ 0.989 0.985 0.991 0.988 0.984 0.991 Bland-Altman analyses typically showed biases of less than one percent
min) with less than eight percent standard deviation of the marker-based
Swing Time (s) 0.920 0.897 0.938 0.898 0.869 0.921
means for all parameters and relatively symmetrical LOAs around
ICC: Interclass correlation coefficients, DLS: Double Limb Support Time. zero. Notably, the results suggest only small systematic error for a subset
of parameters including left stride length (Bias: − 0.004 ± 0.020 m),
ICCs for relative consistency showed nearly identical results except right step cadence (Bias: 0.395 ± 1.952 steps/min), and stride width
for double limb support time calculated using velocity-based gait events. (Bias: 0.001 ± 0.008 s). Given the number of trials analyzed, confidence
Bland-Altman bias was small with narrow limits of agreement (LOA) intervals for the bias between systems were small and therefore the
between systems using velocity-based gait events. The 95% confidence systematic error in these parameters is expected to be negligent.

Table 4
Bland-Altman analysis for all spatiotemporal parameters obtained from marker-based and markerless tracking using velocity-based gait events for calculations. Means
from both systems are provided for context.
Marker Markerless Parameter Bias ± SD CI95 Lower LOA Upper LOA

1.238 ± 0.197 1.238 ± 0.196 Speed (m/s) 0.000 ± 0.020 − 0.002, 0.003 − 0.039 0.039
1.344 ± 0.163 1.343 ± 0.161 Stride Length (m) − 0.002 ± 0.025 − 0.005, 0.002 − 0.050 0.048
0.161 ± 0.039 0.162 ± 0.041 Stride Width (m) 0.001 ± 0.008 0.000, 0.002 − 0.014 0.017
0.672 ± 0.084 0.672 ± 0.083 Step Length (m) − 0.001 ± 0.015 − 0.003, 0.002 − 0.030 0.028
1.097 ± 0.094 1.096 ± 0.095 Cycle Time (s) − 0.001 ± 0.013 − 0.003, 0.001 − 0.027 0.025
0.727 ± 0.075 0.726 ± 0.074 Stance Time (s) − 0.002 ± 0.024 − 0.005, 0.002 − 0.049 0.046
0.369 ± 0.033 0.370 ± 0.028 Swing Time (s) 0.001 ± 0.020 − 0.002, 0.003 − 0.038 0.039
0.548 ± 0.048 0.548 ± 0.049 Step Time (s) − 0.001 ± 0.010 − 0.002, 0.001 − 0.021 0.020
0.358 ± 0.061 0.356 ± 0.057 DLS (s) − 0.002 ± 0.031 − 0.006, 0.002 − 0.063 0.059
55.113 ± 4.636 55.163 ± 4.640 Cadence (strides/min) 0.051 ± 0.809 − 0.058, 0.160 − 1.535 1.636
Left
0.728 ± 0.074 0.728 ± 0.075 Stance Time (s) 0.000 ± 0.020 − 0.003, 0.003 − 0.039 0.039
0.672 ± 0.078 0.671 ± 0.078 Step Length (m) − 0.001 ± 0.015 − 0.003, 0.001 − 0.031 0.029
0.546 ± 0.047 0.547 ± 0.048 Step Time (s) 0.000 ± 0.010 − 0.001, 0.002 − 0.020 0.020
110.599 ± 9.214 110.576 ± 9.320 Step Cadence (steps/min) − 0.022 ± 2.034 − 0.296, 0.251 − 4.009 3.964
1.341 ± 0.165 1.337 ± 0.161 Stride Length (m) − 0.004 ± 0.020 − 0.006, − 0.001 − 0.044 0.036
55.121 ± 4.557 55.158 ± 4.600 Stride Cadence (strides/min) 0.038 ± 0.912 − 0.085, 0.160 − 1.750 1.825
0.368 ± 0.027 0.368 ± 0.026 Swing Time (s) 0.000 ± 0.012 − 0.002, 0.001 − 0.024 0.023
Right
0.726 ± 0.077 0.723 ± 0.074 Stance Time (s) − 0.003 ± 0.028 − 0.007, 0.000 − 0.058 0.052
0.672 ± 0.090 0.672 ± 0.087 Step Length (m) 0.000 ± 0.014 − 0.002, 0.002 − 0.028 0.027
0.550 ± 0.050 0.549 ± 0.049 Step Time (s) − 0.001 ± 0.010 − 0.003, 0.000 − 0.021 0.019
109.975 ± 9.695 110.185 ± 9.536 Step Cadence (steps/min) 0.210 ± 2.015 − 0.061, 0.481 − 3.740 4.160
1.348 ± 0.165 1.349 ± 0.165 Stride Length (m) 0.001 ± 0.029 − 0.003, 0.005 − 0.056 0.059
55.104 ± 4.716 55.168 ± 4.680 Stride Cadence (strides/min) 0.064 ± 0.706 − 0.031, 0.159 − 1.320 1.447
0.371 ± 0.039 0.373 ± 0.030 Swing Time (s) 0.002 ± 0.027 − 0.002, 0.005 − 0.051 0.054

SD: Standard Deviation, CI95: 95% Confidence Interval for Bias, LOA: Limits of Agreement, DLS: Double Limb Support Time.

4
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

Table 5
Bland-Altman analysis for all spatiotemporal parameters obtained from marker-based and markerless tracking using coordinate-based gait events for calculations.
Means from both systems are provided for context.
Marker Markerless Combined Bias ± SD CI95 Lower LOA Upper LOA

1.238 ± 0.197 1.238 ± 0.197 Speed (m/s) 0.000 ± 0.020 − 0.003, 0.002 − 0.039 0.038
1.345 ± 0.163 1.342 ± 0.163 Stride Length (m) − 0.003 ± 0.028 − 0.006, 0.001 − 0.058 0.053
0.163 ± 0.039 0.164 ± 0.041 Stride Width (m) 0.001 ± 0.007 0.000, 0.002 − 0.013 0.016
0.672 ± 0.084 0.671 ± 0.084 Step Length (m) − 0.001 ± 0.017 − 0.003, 0.001 − 0.034 0.031
1.097 ± 0.094 1.095 ± 0.095 Cycle Time (s) − 0.002 ± 0.015 − 0.004, 0.000 − 0.031 0.028
0.728 ± 0.073 0.726 ± 0.075 Stance Time (s) − 0.002 ± 0.019 − 0.005, 0.001 − 0.039 0.036
0.369 ± 0.030 0.369 ± 0.027 Swing Time (s) 0.001 ± 0.014 − 0.002, 0.002 − 0.026 0.027
0.548 ± 0.048 0.548 ± 0.049 Step Time (s) − 0.001 ± 0.011 − 0.003, 0.001 − 0.022 0.021
0.358 ± 0.055 0.358 ± 0.058 DLS (s) 0.000 ± 0.023 − 0.004, 0.003 − 0.046 0.045
55.102 ± 4.628 55.186 ± 4.643 Cadence (strides/min) 0.084 ± 0.849 − 0.030, 0.199 − 1.580 1.749
Left
0.730 ± 0.073 0.728 ± 0.076 Stance Time (s) − 0.002 ± 0.022 − 0.005, 0.001 − 0.045 0.042
0.671 ± 0.077 0.671 ± 0.080 Step Length (m) 0.000 ± 0.017 − 0.002, 0.002 − 0.033 0.032
0.546 ± 0.047 0.547 ± 0.048 Step Time (s) 0.000 ± 0.011 − 0.001, 0.002 − 0.021 0.022
110.671 ± 9.305 110.590 ± 9.282 Step Cadence (steps/min) − 0.082 ± 2.154 − 0.371, 0.208 − 4.304 4.140
1.342 ± 0.165 1.336 ± 0.161 Stride Length (m) − 0.006 ± 0.024 − 0.009, − 0.003 − 0.054 0.042
55.089 ± 4.538 55.209 ± 4.640 Stride Cadence (strides/min) 0.120 ± 0.974 − 0.011, 0.251 − 1.789 2.029
0.367 ± 0.028 0.367 ± 0.025 Swing Time (s) 0.000 ± 0.013 − 0.002, 0.001 − 0.026 0.025
Right
0.725 ± 0.072 0.724 ± 0.074 Stance Time (s) − 0.002 ± 0.016 − 0.004, 0.000 − 0.033 0.029
0.673 ± 0.091 0.671 ± 0.087 Step Length (m) − 0.002 ± 0.016 − 0.004, 0.000 − 0.034 0.029
0.551 ± 0.050 0.549 ± 0.049 Step Time (s) − 0.002 ± 0.011 − 0.004, − 0.001 − 0.023 0.019
109.872 ± 9.574 110.268 ± 9.586 Step Cadence (steps/min) 0.395 ± 1.952 0.133, 0.658 − 3.430 4.220
1.348 ± 0.164 1.349 ± 0.167 Stride Length (m) 0.001 ± 0.032 − 0.003, 0.005 − 0.061 0.063
55.114 ± 4.717 55.163 ± 4.647 Stride Cadence (strides/min) 0.048 ± 0.724 − 0.049, 0.146 − 1.371 1.468
0.371 ± 0.033 0.372 ± 0.029 Swing Time (s) 0.001 ± 0.014 − 0.001, 0.003 − 0.026 0.028

SD: Standard Deviation, CI95: 95% Confidence Interval for Bias, LOA: Limits of Agreement, DLS: Double Limb Support Time.

Compared to previous testing with this markerless system (Ripic For clinically-meaningful differences specific to the groups studied in
et al., 2022), the current results demonstrate greater accuracy for nearly this work, Hass et al. 2014, report small, moderate, and large effect sizes
all spatiotemporal parameters. Specifically, significant improvements for gait speed in PD as 0.06 m•s− 1, 0.14 m•s− 1, and 0.22 m•s− 1 (Hass
were seen for stride width, stance time, swing time, and double limb et al., 2014). LOA in the present results indicate a window of less than
support time. It was previously proposed that the error in spatiotem­ 0.08 m•s− 1 verifying the approach would be sensitive to moderate and
poral parameters between marker-based and markerless systems would large changes in gait speed in PD patients. Differently, the current ac­
be partially due to the gait event method and the location of the prox­ curacy of the markerless approach may be sensitive to changes in the
imal foot segments in the markerless model (Ripic et al., 2022). Stem­ Unified Parkinson’s Disease Rating Scale (UPDRS)-Motor Scale with
ming from this, the latest version of the markerless model included large differences between scale cut points corresponding with 0.09
keypoints at the center of the calcanei, similar to what is available in the m•s− 1 changes in gait speed (Hass et al., 2014). Additionally, Zanardi
marker-based model. It is now currently thought that the error in pre­ and colleagues demonstrated that PD patients may show − 0.17 m•s− 1
vious results was not likely due to gait event methods but was able to be and − 0.16 m differences in gait speed and stride length as compared to
reduced by more closely replicating model definitions and processing age- and sex-matched healthy controls (Zanardi et al., 2021). Both LOAs
methods. Remaining error between systems may be due to the fact that for gait speed and stride length from the present results would indicate
heel marker (keypoint) placements are not perfectly aligned within the that the markerless system would be sensitive to differences in these
foot segments as the markerless keypoint is estimated at the center point variables between PD and healthy age- and sex-matched adults.
of the calcaneus, whereas the marker is placed on the posterior calca­ There are several limitations that may have influenced current re­
neus. This modeling difference can be reduced by defining a virtual sults and the potential impact of the proposed markerless method. The
marker inset in the calcaneus within the marker-based model, whereas use of an over-ground gait testing condition can be challenging when
the optimal location of this virtual marker may be 70–80% along a line implemented across various ages and in PD patients. Choosing the
between the ankle joint center and heel marker. Nevertheless, the small appropriate starting position for each subject and trial is complicated in
distance between placements is expected to have minimal impact on clinical groups and influences the ability to replicate foot falls upon the
walking as shown by the current results. force plate area in the current laboratory configuration. Additionally,
In context to other studies previously compared in (Ripic et al., the objective was to implement gait testing methods which would
2022), the present approach is satisfactory showing similar agreement in replicate subjects’ natural gait pattern and therefore indicating to each
speed, stride length, step length, and cycle time; and better agreement in subject where the foot falls need to be placed may cause further devia­
stride width, step time, stance time, swing time, and double limb support tion from their normal gait patterns. Further, instructing subjects to start
time compared to a similar study (Kanko et al., 2021). LOA in the pre­ each trial with the same leg could also cause abnormal gait patterns. For
sent results are comparable to studies (Bravi et al., 2021; Dubois and these reasons, the starting position was selected to ensure that one gait
Bresciani, 2018; Geerse et al., 2015; Kanko et al., 2021; Yeo and Park, cycle would occur within the testing areas for both systems without
2020) that evaluated the validity of spatiotemporal variables to other requiring that foot falls occur overtop the force plates and no in­
benchmark systems. For spatial variables, LOA in the present results structions were given to start with the same leg for each trial. This is
were relatively symmetrical around zero and in range of previous study limiting in that kinematic gait event methods were not compared to
results for both gait event methods. This was also the case for temporal force plate-derived gait methods, although full force plate-derived gait
variables, with double limb support time showing improvements events cannot be achieved with only two force plates in the current
compared to previous laboratory results and those using a similar laboratory configuration. Our results are dependent upon the validity of
markerless-based method (Kanko et al., 2021). kinematically-determined gait events which were previously shown to

5
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

Fig. 1. Example Bland-Altman plot for gait speed. 95% confidence intervals for bias, lower and upper limits of agreement (LoA) are indicated by the shaded region.

have 16–33 ms error across healthy participants and in those with variables should be re-evaluated against full force plate or pressure mat
neurological pathologies (Zeni et al., 2008). Further studies addressing systems to understand the effect of kinematically-determined gait
the validity of spatiotemporal parameters derived from KinaTrax mar­ events.
kerless tracking may strengthen current results with the inclusion of
comparisons to either marker-based systems combined with force CRediT authorship contribution statement
sensing treadmills, pressure mats, or foot-switch systems.
In conclusion, the present study aimed to evaluate the validity of Zachary Ripic: Validation, Methodology, Investigation, Formal
spatiotemporal parameters obtained from markerless tracking. Results analysis, Data curation, Conceptualization. Joseph F. Signorile:
show improvements in agreement between marker-based and marker­ Writing – review & editing, Writing – original draft, Formal analysis.
less systems compared to previous results from the laboratory. This may Thomas M. Best: Investigation, Conceptualization. Kevin A. Jacobs:
be largely due to the inclusion of additional keypoints in the markerless Investigation, Conceptualization. Mitch Nienhuis: Project administra­
model which more closely approximates the marker-based model used. tion, Data curation. Cole Whitelaw: Project administration, Data
This methodological factor brought results in range with other studies curation. Caden Moenning: Data curation, Project administration.
that compared spatiotemporal parameters from various systems to Moataz Eltoukhy: Writing – original draft, Validation, Supervision,
marker-based systems or pressure mats. Gait velocity and stride length Project administration, Methodology, Investigation, Data curation,
LOAs show that the markerless method would be sensitive enough to Conceptualization.
distinguish differences in PD and healthy subjects, and in some regard,
differences within PD groups. Further work should seek to reduce errors
in spatiotemporal parameters to confidently distinguish between Declaration of Competing Interest
UPDRS-Motor Scale. To do this, offsets between heel keypoints and heel
marker placements should be explored, and validity of spatiotemporal The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence

6
Z. Ripic et al. Journal of Biomechanics 155 (2023) 111645

the work reported in this paper. Lu, T.-W., Chang, C.-F., 2012. Biomechanics of human movement and its clinical
applications. Kaohsiung J. Med. Sci. 28 (2), S13–S25.
Mehdizadeh, S., Nabavi, H., Sabo, A., Arora, T., Iaboni, A., Taati, B., 2021. Concurrent
Appendix A. Supplementary data validity of human pose tracking in video for measuring gait parameters in older
adults: a preliminary analysis with multiple trackers, viewing angles, and walking
Supplementary data to this article can be found online at https://doi. directions. J. NeuroEngineering Rehabilitation 18, 139.
Moro, M., Marchesi, G., Hesse, F., Odone, F., Casadio, M., 2022. Markerless vs. Marker-
org/10.1016/j.jbiomech.2023.111645. Based Gait Analysis: A Proof of Concept Study. Sensors 22 (5), 2011.
Mündermann, L., Corazza, S., Andriacchi, T.P., 2006. The evolution of methods for the
References capture of human movement leading to markerless motion capture for
biomechanical applications. J. Neuroengineering Rehabilitation 3 (1).
Newell, A., Yang, K., Deng, J., Year Stacked Hourglass Networks for Human Pose
Bravi, M., Massaroni, C., Santacaterina, F., Di Tocco, J., Schena, E., Sterzi, S., Bressi, F., Estimation. In ECCV.
Miccinilli, S., 2021. Validity Analysis of WalkerView(TM). Sensors 21 (14), 4795. Oh, J., Eltoukhy, M., Kuenze, C., Andersen, M.S., Signorile, J.F., 2020. Comparison of
Desmarais, Y., Mottet, D., Slangen, P., Montesinos, P., 2021. A review of 3D human pose predicted kinetic variables between Parkinson’s disease patients and healthy age-
estimation algorithms for markerless motion capture. Comput. Vis. Image Underst. matched control using a depth sensor-driven full-body musculoskeletal model. Gait
212, 103275. & Posture 76, 151–156.
Donno, L., Sansone, V., Galluzzo, A., Frigo, C.A., 2022. Walking in the Absence of Razavian, R.S., Greenberg, S., McPhee, J.J.E.o.B.E., 2019. Biomechanics imaging and
Anterior Cruciate Ligament: The Role of the Quadriceps and Hamstrings. 12, 8667. analysis. 488-500.
Dubois, A., Bresciani, J.P., 2018. Validation of an ambient system for the measurement of Ren, S., He, K., Girshick, R.B., Sun, J., 2015. Faster R-CNN: Towards Real-Time Object
gait parameters. J Biomech 69, 175–180. Detection with Region Proposal Networks. 39, 1137-1149.
Eltoukhy, M., Kuenze, C., Andersen, M.S., Oh, J., Signorile, J., 2017. Prediction of ground Ripic, Z., Signorile, J.F., Kuenze, C., Eltoukhy, M., 2022. Concurrent validity of artificial
reaction forces for Parkinson’s disease patients using a kinect-driven musculoskeletal intelligence-based markerless motion capture for over-ground gait analysis: A study
gait analysis model. Med. Eng. Phys. 50, 75–82. of spatiotemporal parameters. J. Biomech. 143, 111278.
Geerse, D.J., Coolen, B.H., Roerdink, M., Fasano, A., 2015. Kinematic Validation of a Sandler, M., Howard, A.G., Zhu, M., Zhmoginov, A., Chen, L.-C.J.I.C.C.o.C.V.,
Multi-Kinect v2 Instrumented 10-Meter Walkway for Quantitative Gait Assessments. Recognition, P., 2018. MobileNetV2: Inverted Residuals and Linear Bottlenecks.
PLoS One 10 (10), e0139913. 4510-4520.
Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep Learning. Massachusetts Institute of Shelhamer, E., Long, J., Darrell, T., 2017. Fully Convolutional Networks for Semantic
Technology, Cambridge, MA. Segmentation. 39, 640-651.
Guffanti, D., Brunete, A., Hernando, M., Rueda, J., Navarro Cabello, E., 2020. The Tan, M., Le, Q., 2019. EfficientNet: Rethinking Model Scaling for Convolutional Neural
Accuracy of the Microsoft Kinect V2 Sensor for Human Gait Analysis. Sensors 20 Networks, in: Kamalika, C., Ruslan, S. (Eds.), Proceedings of the 36th International
(16), 4405. Conference on Machine Learning. PMLR, Proceedings of Machine Learning Research,
Hass, C.J., Bishop, M., Moscovich, M., Stegemöller, E.L., Skinner, J., Malaty, I.A., Wagle pp. 6105–6114.
Shukla, A., McFarland, N., Okun, M.S., 2014. Defining the clinically meaningful Thomann, A.E., Berres, M., Goettel, N., Steiner, L.A., Monsch, A.U., 2020. Enhanced
difference in gait speed in persons with Parkinson disease. J. Neurol. Phys. Ther. 38, diagnostic accuracy for neurocognitive disorders: a revised cut-off approach for the
233–238. Montreal Cognitive Assessment. Alzheimer’s Res. Therapy 12, 39.
He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep Residual Learning for Image Recognition. Topley, M., Richards, J.G., 2020. A comparison of currently available optoelectronic
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016, motion capture systems. J. Biomech. 106, 109820.
770–778. Winter, D.A. (Ed.), 2009. Biomechanics and Motor Control of Human Movement. Wiley.
Huang, G., Liu, Z., Maaten, L.V.D., Weinberger, K.Q., Year Densely Connected Yeo, S.S., Park, G.Y., 2020. Accuracy Verification of Spatio-Temporal and Kinematic
Convolutional Networks. In 2017 IEEE Conference on Computer Vision and Pattern Parameters for Gait Using Inertial Measurement Unit System. Sensors 20 (5), 1343.
Recognition (CVPR). Yeung, L.-F., Yang, Z., Cheng, K.-C.-C., Du, D., Tong, R.-K.-Y., 2021. Effects of camera
Kanko, R.M., Laende, E.K., Strutzenberger, G., Brown, M., Selbie, W.S., DePaul, V., viewing angles on tracking kinematic gait patterns using Azure Kinect, Kinect v2 and
Scott, S.H., Deluzio, K.J., 2021. Assessment of spatiotemporal gait parameters using Orbbec Astra Pro v2. Gait & Posture 87, 19–26.
a deep learning algorithm-based markerless motion capture system. J. Biomech. 122, Zago, M., Sforza, C., Pacifici, I., Cimolin, V., Camerota, F., Celletti, C., Condoluci, C., De
110414. Pandis, M.F., Galli, M., 2018. Gait evaluation using inertial measurement units in
Koo, T.K., Li, M.Y., 2016. A Guideline of Selecting and Reporting Intraclass Correlation subjects with Parkinson’s disease. J. Electromyogr. Kinesiol. 42, 44–48.
Coefficients for Reliability Research. J. Chiropr. Med. 15 (2), 155–163. Zahradka, N., Verma, K., Behboodi, A., Bodt, B., Wright, H., Lee, S.C.K., 2020. An
Krizhevsky, A., Sutskever, I., Hinton, G.E.J.C.o.t.A., 2012. ImageNet classification with Evaluation of Three Kinematic Methods for Gait Event Detection Compared to the
deep convolutional neural networks. 60, 84-90. Kinetic-Based ‘Gold Standard’. Sensors 20 (18), 5272.
Kumarasiri, R., Niroshan, A., Lantra, Z., Madusanka, T., Edussooriya, C.U.S., Rodrigo, R., Zanardi, A.P.J., da Silva, E.S., Costa, R.R., Passos-Monteiro, E., dos Santos, I.O., Kruel, L.
Year Gait Analysis Using RGBD Sensors. In 2018 15th International Conference on F.M., Peyré-Tartaruga, L.A., 2021. Gait parameters of Parkinson’s disease compared
Control, Automation, Robotics and Vision (ICARCV). with healthy controls: a systematic review and meta-analysis. Sci. Rep. 11, 752.
Lonini, L., Moon, Y., Embry, K., Cotton, R.J., McKenzie, K., Jenz, S., Jayaraman, A., Zeni, J.A., Richards, J.G., Higginson, J.S., 2008. Two simple methods for determining
2022. Video-Based Pose Estimation for Gait Analysis in Stroke Survivors during gait events during treadmill and overground walking using kinematic data. Gait
Clinical Assessments: A Proof-of-Concept Study. Digital Biomarkers 6, 9–18. Posture 27 (4), 710–714.