Comparison of concurrent and asynchronous running kinematics and kinetics from marker-based motion capture and markerless motion capture under two clothing conditions

As markerless motion capture is increasingly used to measure three-dimensional human pose, it is important to understand how markerless results can be interpreted alongside historical marker-based data and how they are impacted by clothing. We compared concurrent running kinematics and kinetics between marker-based and markerless motion capture, and between two markerless clothing conditions. Thirty adults ran on an instrumented treadmill wearing motion capture clothing while concurrent marker-based and markerless data were recorded, and ran a second time wearing athletic clothing (shorts and t-shirt) while markerless data were recorded. Differences calculated between the concurrent signals from both systems, and also between each participant’s mean signals from both asynchronous clothing conditions were summarized across all participants using root-mean-square differences. Most kinematic and kinetic signals were visually consistent between systems and markerless clothing conditions. Between systems, joint center positions differed by 3 cm or less, sagittal plane joint angles differed by 5° or less, and frontal and transverse plane angles differed by 5°-10°. Joint moments differed by 0.3 Nm/kg or less between systems. Differences were sensitive to segment coordinate system definitions, highlighting the effects of these definitions when comparing against historical data or other motion capture modalities.


Introduction 55
Motion capture has long been used to quantify movement patterns, monitor interventions and has 56 potential as a tool for injury prediction. The latter has been a particular focus for those interested 57 in its application to running biomechanics. However, due to the inherent time, cost, environment, 58 and expertise constraints associated with traditional marker-based motion capture approaches, 59 the successes of this tool have been limited. Likewise, wearable sensors such as inertial 60 measurement units (IMUs) have limitations, specifically related to magnetic field effects such as 61 those generated by laboratory treadmills 1,2 and are best suited to reporting quantities such as 62 accelerations and angular rates which they measure directly, rather than estimating 63 intersegmental measures such as joint angles. 64 As emerging technology, markerless motion capture has the potential to capture high 65 resolution three dimension (3D) biomechanical signals of running in a more accessible system 66 that avoids potential effects of applying markers or sensors. Ongoing developments, validations, 67 and applications demonstrate the promise of this form of motion capture technology for increasing 68 the usefulness of motion capture-based running gait analysis. Markerless motion capture systems 69 such as Theia3D (Theia Markerless Inc., Kingston, Ontario, Canada) can measure kinematics 70 using multiple synchronized video cameras with few restrictions on the data collection 71 environment, subject attire, and a significantly reduced collection time compared to marker-based 72 motion capture. Markerless systems have demonstrated success in measuring comparable 73 kinematic signals to those from marker-based motion capture during walking 3,4 , boxing 5 , and 74 running 6 . While marker-based motion capture may not provide a true gold standard for 75 biomechanical data collection due to issues with soft tissue artefact and marker placement 76 variation, it is a historically accepted technology and forms the basis of much of the running 77 biomechanics literature. Understanding how data collected with markerless motion capture 78 compares to marker-based data, as well as the factors contributing to any differences between 79 Marker-based and markerless motion capture data were analyzed separately in Visual3D. 144 Analog signals (force data) from the instrumented treadmill were imported into the markerless 145 MoCap clothing and Sport clothing .c3d files using the Visual3D 'Import_Signals_From_C3D_File' 146 pipeline command, allowing these signals to be used with the markerless kinematics to obtain 147 joint kinetics and create force-based gait events. Analog signals were filtered using a low-pass 148 critically-damped filter with a cut-off frequency of 20 Hz to match the kinematic filter cut-off 149 frequency 9 , and the Visual3D 'FP Auto Baseline' pipeline command was used to clean the force 150 platform signals. A vertical ground reaction force threshold of 50 N was used to detect foot contact 151 events. 152 Kinematic and kinetic signals including joint positions, segment angles, joint angles, and 153 internal net joint moments were calculated in Visual3D, and were time-normalized to 101 data 154 points, representing one complete running gait cycle (0-100%) from heel strike to ipsilateral heel 155 strike. Segment angles were calculated using a XYZ Cardan sequence relative to the global 156 coordinate system and were included since they capture the 3D orientation of individual body 157 segments. Joint angles were calculated using an XYZ Cardan sequence for all joints except the 158 ankles, for which a XZY Cardan sequence was used. Joint angles were included since they 159 capture the relative orientation of adjacent segments and are kinematic measures of interest for 160 most users of motion capture. Joint moments were normalized to participant mass and resolved 161 in the proximal segment. All data were analyzed further using python (version 3.10.6). 162 Four participants' markerless Sport clothing trial videos had an inconsistent number of 163 frames between cameras, preventing their processing in Theia3D; these participants were 164 excluded from the kinematic comparisons, which then included 26 participants (12f/14m, mean 165 (SD) age: 23.5 (4.0) years, height: 1.8 (0.1) m, mass: 69.9 (9.9) kg, BMI: 22.8 (2.9) kg/m 2 ). Six 166 participants' center-of-pressure data from the instrumented treadmill were found to be erroneous, 167 one of which was also part of the previously described kinematic exclusion group. The kinetic 168 comparison sample therefore included 21 participants (10f/11m, mean (SD) age: 23.3 (3.8) years, 169 height: 1.8 (0.1) m, mass: 69.9 (11.0) kg, BMI: 22.8 (2.8) kg/m 2 ). Running gait cycles with marker 170 tracking issues were excluded. For comparison between the markerless and marker-based 171 motion capture, only matched gait cycles were used. 172 Differences between signals from corresponding running gait cycles for the concurrent marker-173 based and markerless MoCap clothing datasets were calculated and summarized across all 174 participants as root-mean-square differences (RMSD). Participants' mean signals for the 175 asynchronous markerless MoCap clothing and Sport clothing conditions were used to assess 176 differences between markerless clothing conditions and were summarized across all 177 participants as RMSDs. The inter-cycle variability of the signals from each dataset was 178 calculated 10 , and the inter-condition variability of markerless signals was also calculated to 179 examine the reliability of markerless measurements across different clothing conditions (similar 180 to 'inter-session' variability 10 ). 181 182

Alternative Marker-Based Modelling 183
To investigate the effect of marker-based model definitions on kinematic and kinetic 184 comparisons, results were calculated for two additional models. The first alternative model was 185 our previous standard marker-based model that was used in prior publications 3 (Supplementary 186 Materials, Table A). The second alternative model used the same set of markers but applied 187 atypical segment definitions so that the resulting kinematics best aligned with the markerless 188 results (Supplementary Materials, Table B). The modifications made to arrive at this model 189 definition were selected on the basis of improving the agreement between the marker-based 190 and markerless signals, and serve a purely illustrative purpose. This model was created by 191 modifying the primary model used in the analysis, using techniques such as virtual markers that 192 are offset relative to their real counterparts to define segment axes.  across all three datasets, and had mean RMSDs of 5.2° or lower (Table 2). Ankle toe-in/toe-out 235 patterns were also somewhat visually consistent between all three datasets during the swing 236 phase, and had a mean RMSD of 8.0°. The hip internal/external rotation, knee ab/adduction and 237 internal/external rotation, and ankle inversion/eversion angles captured different patterns and 238 levels of variability between motion capture systems but were consistent between both markerless 239 clothing conditions (mean RMSDs: 5-10°). The mean RMSD values for the asynchronous 240 markerless data (MoCap clothing, Sport clothing) were all smaller than those for the concurrent 241 marker-based and markerless data, indicating that the differences between markerless joint angle 242 measures from different trials with different participant clothing were smaller than the differences 243 between joint angles from marker-based and markerless systems from the same running gait 244 cycles. 245   (Table 3). The average inter-cycle variability across all joint angles was slightly 256 smaller for the marker-based data (1.3°) compared to those for both markerless datasets (1.5°). 257 The inter-condition variability for joint angles between markerless conditions ranged from 1.3°-258 2.8° (mean 2.0°), leading to variability ratios of 1.2-1.7 (mean 1.3). 259

Kinetics 265
Lower limb joint moments from all three datasets captured visually similar patterns for 266 most joint moments, with differences observed in the knee flexion/extension and ab/adduction 267 moments, ankle toe-in/toe-out moments, and large differences observed for the ankle 268 inversion/eversion moments (Figure 5). The marker-based knee moments had greater peak 269 magnitudes and greater variability compared to both markerless conditions in all three planes. 270 The ankle inversion/eversion moments measured by the marker-based system were inversion 271 moments across all subjects and had considerably greater variability compared to those 272 measured by the markerless system, which were primarily eversion moments across all subjects 273 in both conditions. Additionally, the ankle toe-in/toe-out joint moments measured by the marker-

Alternative Marker-Based Modelling Results 285
The results of this study when two alternative marker-based models were applied are 286 included in the Supplementary Materials (Tables A, B; Figures A-F). The results from the model 287 used in previous studies 3 were consistent with the results of that study, with slight variations due 288 to the differences in task (walking versus running) (Supplementary Materials, Figures A, B, C). 289 The results from the modified version of the model used here showed the same or greater 290 similarity between marker-based and markerless systems for all segment angles, joint angles, 291 and joint moments (Supplementary Materials, Figures D, E, F). 292

Discussion 293
The purpose of this study was to compare kinematic and kinetic measures of treadmill running 294 from concurrent marker-based and markerless motion capture, and between asynchronous 295 markerless motion capture under varying clothing conditions. Overall, we found comparable 296 results between the simultaneous marker-based and markerless motion capture data and 297 between clothing conditions for the markerless system. 298 There are several limitations of this study. Challenges with data collection resulted in some 299 lost participants, but without a significant change in the demographics of the study population. 300 The self-selected treadmill running speed of this group is lower than typical over-ground running 301 speeds, which is likely due to previously observed perception differences in treadmill and over-302 ground running, and is consistent with previous findings 11,12 . Likewise, the framerate of 85 Hz for 303 data collection is lower than is usually used in running studies but was a limitation of the hardware 304 to allow maximum image resolution and synchronization across both motion capture systems. 305 Although this is a relatively lower frame rate than is typically used to capture running, the purpose 306 was to compare motion capture modalities rather than to study running biomechanics. Using 307 different markerless motion capture hardware or capturing lower resolution videos would allow 308 higher frequency data to be collected. We would anticipate that a higher running speed recorded 309 at a higher framerate would have comparable results to those presented here. The order of the 310 two clothing conditions was divided evenly across all participants, but not randomized and not 311 considered in the data analysis. However, this would have no impact on the comparison of 312 simultaneous marker-based and markerless data, and the lack of differences observed between 313 clothing conditions implies that this was not an issue. Furthermore, the precise segment 314 coordinate system definitions used by Theia3D are not publicly available, so our marker-based 315 definitions that are intended to mimic the markerless segment definitions could possibly be 316 improved upon; but in any case, the anatomical landmarks used by the markerless system are 317 also not available and likely differ slightly in position from equivalent marker positions by nature 318 of how they are created and tracked. 319 Joint position differences between marker-based and markerless concurrent data were 320 found to be 3 cm or smaller for all joints, which is consistent with previous comparisons of these 321 systems during treadmill walking 3 . Joint position component differences were also mostly 322 consistent with those of Tang et al., however they did not report 3D Euclidean distances 6 . 323 Segment and joint angles demonstrated similar agreement between motion capture 324 systems as in our previous study for treadmill walking gait 3 , and excellent agreement across 325 asynchronous markerless clothing conditions. Sagittal plane joint angles between motion capture 326 systems had mean RMSD of 5.2° or less for all three joints, consistent with the findings of Tang 327 et al. 6 . The differences between systems measured as mean RMSD over the entire gait cycle are 328 comparable to reported minimal detectable change (MDC) values for marker-based motion 329 capture of treadmill running at initial contact 13 . Compared to the marker-based MDC values, the 330 RMSD values between systems are 3.1° smaller (for ankle dorsiflexion) to 4.7° greater (for knee 331 ab/adduction). The between-system agreement observed here is also consistent with previous 332 comparisons of these systems during treadmill walking 3 , which showed larger differences for 333 segment angles about the segment longitudinal axes, and for frontal and transverse plane angles. 334 Kinematic and kinetic signals were highly consistent between the markerless MoCap 335 clothing and Sport clothing conditions, except for pelvis segment angle x-and z-components. 336 Interpreting kinematic differences between the two markerless motion capture clothing conditions 337 is made challenging by the fact that they may result from actual changes in movement patterns 338 between the two running trials or due to differences in markerless keypoint detections as a result 339 of the change in clothing. The former was indeed the case for at least one participant, whose 340 markerless joint angles were considerably different between the MoCap and Sport clothing 341 conditions, and were confirmed from raw video data to have visibly greater ranges of motion in 342 hip, knee, and ankle flexion for the Sport clothing condition (Supplementary Material, Figure H). 343 The RMSD values between clothing conditions, which could be considered analogous to a 344 participant returning on a second day in different clothing, were smaller than published MDC 345 values at initial contact 13 , with the exception of hip flexion (0.2° greater). These findings are 346 consistent with a running repeatability study over three repeated visits at varying running speeds 347 using the same markerless motion capture software 14 . Their findings and ours support the use of 348 markerless motion capture methods to overcome the limitations of marker placement variability 349 with marker-based motion capture, which has been a challenge for longitudinal running studies 15 . 350 Joint angle inter-cycle variability was small and similar across the marker-based data 351 (1.3°) and both markerless datasets (1.5°). Inter-cycle variability is made up of natural stride-to-352 stride variation in subject biomechanics, as well as measurement variability inherent to the motion 353 capture systems. Inter-condition variability across markerless clothing conditions was larger than 354 the inter-cycle variabilities, and presented an average increase of 30% in measurement variability. 355 The inter-condition variability is made up of inter-cycle variability as described above, plus natural 356 inter-trial variability, measurement variability due to the varying participant clothing, and 357 measurement variability inherent to the motion capture system. The inter-condition variability was 358 smaller than the mean RMSD between concurrent motion capture systems for all joint angles, 359 indicating that differences between markerless clothing conditions remained smaller than 360 differences between motion capture systems. 361 Joint moments showed similarity in pattern and magnitude between systems and across 362 clothing conditions for the ankle, knee, and hip sagittal plane moments, hip ab/adduction moment, 363 and hip and knee internal/external rotation moments, while similar patterns of different magnitude 364 were observed for knee ab/adduction and ankle toe-in/toe-out moments. However, the ankle 365 inversion/eversion moments displayed different patterns, with the marker-based moments being 366 mostly inversion moments while the markerless moments were mostly eversion moments. 367 Differences in sagittal plane joint moment peak magnitudes between marker-based and 368 markerless joint moments were mostly consistent with the findings of Tang et al., but they did not 369 report frontal or transverse plane joint moments for either system 6 . The differences in ankle 370 inversion/eversion and toe-in/toe-out moments between systems can be explained by differences 371 in the foot and shank segment orientations, as the selection and orientation of the coordinate 372 system in which joint moments are resolved can entirely change their direction 16  an optimally-modified model for matching marker-based and markerless signals, and it is entirely 393 possible that different modifications could be implemented to greater success in matching the 394 signals between systems. We advocate for the use of the automated coordinate system definition 395 from Theia3D to facilitate standardization for multi-centre and longitudinal research, but found this 396 a useful exercise to understand potential differences in how data has been collected from our 397 historical marker-based approaches. This demonstrated the sensitivity of concurrent comparisons 398 to these factors, which is an important point that should be understood when using and evaluating 399 different motion capture solutions. These results highlight the challenges in combining data 400 collected under different model definitions, and the potential benefit in using markerless motion 401 capture to provide standardized results that can be pooled from multiple sites. 402 The results of this study indicate that Theia3D can measure treadmill running kinematics 403 and kinetics with minimal effects from differing attire, and which are comparable to measures from 404 marker-based motion capture. This study reiterates the influence of segment definition on 405 resulting biomechanical signals. Careful consideration of these effects should be taken when 406 conducting comparisons between measurement systems and relative to historical data that 407 utilized potentially different segment definitions. 408      MoCap clothing), using the marker-based model defined in Table B to better match the results from   495 Theia3D for demonstration purposes.  Table B to better match the results from Theia3D for 501 demonstration purposes. MoCap clothing), using the marker-based model defined in Table B to better match the results from 507 Theia3D for demonstration purposes. 508

Example individual participant kinematic waveforms 509
Raw joint angle data for a representative participant with consistent running biomechanics 510 (Figure G) and a participant with altered running biomechanics when wearing MoCap clothing 511 compared to their self-selected Sport clothing ( Figure H). 512