Development, evaluation and application of a novel markerless motion analysis system to understand push-start technique in elite skeleton athletes.

This study describes the development, evaluation and application of a computer vision and deep learning system capable of capturing sprinting and skeleton push start step characteristics and mass centre velocities (sled and athlete). Movement data were captured concurrently by a marker-based motion capture system and a custom markerless system. High levels of agreement were found between systems, particularly for spatial based variables (step length error 0.001 ± 0.012 m) while errors for temporal variables (ground contact time and flight time) were on average within ± 1.5 frames of the criterion measures. Comparisons of sprinting and pushing revealed decreased mass centre velocities as a result of pushing the sled but step characteristics were comparable to sprinting when aligned as a function of step velocity. There were large asymmetries between the inside and outside leg during pushing (e.g. 0.22 m mean step length asymmetry) which were not present during sprinting (0.01 m step length asymmetry). The observed asymmetries suggested that force production capabilities during ground contact were compromised for the outside leg. The computer vision based methods tested in this research provide a viable alternative to marker-based motion capture systems. Furthermore, they can be deployed into challenging, real world environments to non-invasively capture data where traditional approaches are infeasible.

Introduction

The skeleton push start requires the athlete to accelerate a sled to a high velocity before loading onto it to adopt a prone driving position. A fast start is considered important to overall success [1, 2] where contributions to successful skeleton push start performance include attaining a high pre-load velocity and executing an effective loading phase [3]. Physical attributes, such as lower limb power and sprinting ability, explain a large portion of the variance in the sled velocity attained [4] and as such the prescription of regular sprint training is often used as physical preparation modality. However, Colyer et al. [4] also noted that factors related to technique are likely to account for some of the remaining variance in performance.

Improved sprint times in elite skeleton athletes have been associated with increases in loading distance and pre-load velocity [5]. However, the enhancement of sprinting ability did not necessarily guarantee a faster start as reductions in loading effectiveness have been observed under higher velocity loading conditions [5]. It appears, therefore, that while sprint training has the potential to enhance push start performance, the relationship between these two activities is a complex one and pushing should not simply be viewed as bent over sprinting. Comparisons of kinematics between training activities and their target skill has demonstrated that similar movement patterns can facilitate task specific adaptations [6] and via transfer of training, enhanced performance of the target skill [1, 7]. Therefore, capturing the underlying kinematics of the athlete during pushing and sprinting is required to further understand the transfer of sprinting to pushing a skeleton sled. Such information would allow for the comparison of pushing kinematics to the more researched activity of sprinting and provide practitioners with a conceptual understanding to guide the selection of skeleton specific training activities [8].

To date, limited push start kinematic information is available, perhaps due in part to the limitations associated with current motion capture technology and the challenging environment that skeleton takes place in. For example, dynamic outdoor lighting conditions are challenging to control and account for during motion capture with marker-based systems. Furthermore, placement of reflective markers is both time consuming and invasive with the potential to alter technique and thus reduce ecological validity. Advances in computer vision and deep learning are revolutionising all aspects of movement sciences, providing viable alternatives to the traditional marker-based motion capture technologies that are considered the gold standard for many biomechanical research applications [9]. Computer vision and deep learning based approaches have advantages over marker-based motion capture and inertial measurement units (IMUs) in that they are capable of fully automating the capture and extraction of information from an image or sequence of images in a manner that is completely non-invasive to the participant. Furthermore, by capturing and storing regular images data can easily be reprocessed with the latest algorithms for improved results in this rapidly developing field. Deep convolutional neural networks (CNN) have been highly influential in the computer vision research and application, excelling at tasks such as segmentation (e.g. segmenting the outline of a person in an image) and key-point estimation (e.g. estimation of joint centre coordinates). While such methods are computationally intensive and require large amounts of training data, they consistently outperform more traditional methods [10] such as the use of binary image operations or shallow, fully connected neural networks.

Previous work from our research group presented a computer vision based multi-camera system capable of non-invasively capturing accurate step characteristic data during over-ground running [11]. Validation against force plates and manual camera-based analysis showed temporal variables to fall within 1.5 frames of the criterion data while spatial variables demonstrated errors of < 1 cm. Needham et al. [12] provided modifications to this system incorporating deep learning based methods in order to robustly capture step characteristic data during the skeleton push start at an outdoor training facility. Additionally, the proposed system included athlete and sled mass centre velocities averaged across each step. Evans et al. [13], further enhanced the tracking of athlete and sled providing accurate velocities as a function of time. In the work presented here, we build upon the previously proposed systems to present, validate and apply a complete motion analysis system that can capture step characteristics and mass centre behaviour for both sprinting and skeleton push starts. Furthermore, the complete system described here offers notable performance increases over previous work. Therefore, the aim of this study was to:

Compare the performance of a computer vision and deep learning based system to that of a traditional marker-based motion capture system in a challenging real-world environment (skeleton push-track).

Utilise the proposed system to compare biomechanical indicators of performance (step characteristics and mass centre velocities) during the skeleton push start and regular acceleration sprinting.

Materials and methods

Twelve international skeleton athletes (seven males [1.81 ± 0.05 m, 83.37 ± 2.73 kg], five females [1.71 ± 0.03 m, 70.04 ± 1.44 kg]) were recruited. All participants provided written informed consent. Ethical approval was provided by the University of Bath’s research ethics committee. Approval number—EP 18/19 052. Each athlete attended two testing sessions. During session one, each athlete performed three maximal effort dry-land push starts at the University of Bath’s outdoor push track. During session two, each athlete performed two 10 m sprints, two 15 m sprints and two 20 m sprints. After each pair of sprints, the starting line was moved backwards to allow different parts of the acceleration phase to be captured (Fig 1).

Fig 1

Schematic of the experimental set-up at the push-track (top) and on the indoor track (bottom).

The marker-based and markerless cameras are denoted in black and green, respectively. The capture volume at the push-track was ~10 m long and was centred ~10-m from the block (red dashed area). On the indoor track, a ~10 m volume was captured across multiple trials with three different starting positions. This allowed a 0–30 m section to be reconstructed (red dashed area).

During both sessions, motion data were captured concurrently using two motion capture systems. Criterion data were captured using a 15-camera marker-based motion capture system (Oqus, Qualisys AB, Gothenburg, Sweden) while additional data were captured using a custom 9-camera (1920 × 1080 pixel resolution, ~90° field of view, JAI sp5000c, JAI ltd, Denmark) computer vision system. During session one, both camera systems were positioned around the push track in order to capture the pushing action between 5 m and 15 m from the starting block (Fig 1) where there was a constant gradient of ~2%. During session two, both cameras systems were positioned around an indoor sprints tracks in order to capture the sprinting technique between 0–10 m, 10–20 m and 20–30 m from a three-point start.

Motion capture systems were time-synchronised by means of a periodic TTL-pulse generated by the custom system’s master frame grabber to achieve a frame locked sampling frequency of 200 Hz in both systems. Additionally, start and stop trigger signals for both systems were generated by the master frame grabber on the computer vision system. This ensured that not only did both camera systems start and stop at the same time but that frames were captured by all cameras in unison without drift.

The Qualisys system was calibrated as per the manufacture’s specifications. The custom camera system used a binary dot matrix to initialise each camera’s intrinsic [14] before extrinsic parameters were solved via Sparse Bundle Adjustment [15] to provide a globally optimal calibration. A right-handed coordinate system was defined for both systems by placing a Qualisys L-Frame in the centre of the push track, 10 m from the start block. In order to refine the alignment of each system’s Euclidean space, a single marker was moved randomly through the capture volume and tracked by both systems. This marker data provided points with which the spatial alignment could be optimised in a least-squares sense. To assess the reconstruction accuracy of both systems a wand was moved through the capture volume and tracked by both systems before the mean (± SD) resultant vector magnitude was computed and compared to the known dimensions of the wand.

To capture criterion data a full body marker set comprising of 44 individual markers and four clusters were attached to each participant to create a full body six degrees of freedom (6DoF) model (bilateral feet, shanks and thighs, pelvis and thorax, and bilateral upper arms lower arms, and hands). Four additional markers were placed on the sled to its track position and orientation. Following labelling and gap filling of trajectories (Qualisys Track Manager v2019.3, Qualisys, Gothenburg, Sweden) data were exported to Visual 3D (v6, C-Motion Inc, Germantown, USA) where raw trajectories were low-pass filtered (Butterworth 4th order, cut-off 12 Hz) and a 6DoF inverse kinematics (IK) constrained model was computed. Athlete mass centres (CoM) were computed using the model described by de Leva [16]. Additionally, the sled was modelled as a rigid object with uniformly distributed mass. Filtered marker trajectories and mass centre locations were exported to a custom Python script (v3.7, Python Software Foundation, USA) where mass centre derivatives were computed using a finite central differences method and touch-down (TD) and toe-off (TO) events were computed according to the method described by Handsaker et al. [17]. Computing TD and TO events permitted the calculation of step characteristics including step length (SL), step frequency (SF), step velocity (SV), step time (ST), ground contact time (GCT) and flight time (FT).

This research builds upon the computer vision based foot contact detection algorithms presented by Evans et al. [11] and Needham et al. [12]. Challenging lighting conditions at the push-track caused severe problems for typical background subtraction methods, preventing robust segmentations of the athlete (foreground mask) from the background. As such, traditional background subtraction was replaced with the CDCL-human-part-segmentation [18] and DensePose [19] which both implement a convolutional neural network (CNN) to detect and segment body parts in an image (S1 Fig in S1 File). This approach proved robust to dynamic lighting conditions and had the added advantage of not segmenting the sled as foreground. To detect approximate footfall locations the scene was divided into a horizontal grid of cells, and the occupancy of each cell at various heights calculated by accumulating the proportion of the projection that is foreground in each camera view. Peaks in the ground plane occupancy map larger than a threshold represent cells that are highly likely to contain the on-ground foot. Peaks at knee and body height can be used to verify those ground-plane peaks. Approximate timing and location of contact events can be determined by monitoring occupancy over time. Setting the ground plane to be 0.025 m above the actual ground helped to avoid partial occlusions created by the sled (S2 and S3 Figs in S1 File).

Foot position was further refined by initialising an approximately foot-sized 3D bounding box along the axis representing the direction of travel. The position of the bounding box was initialised using foot contact information from the ground plane occupancy maps before the position was optimised to fit the foot. Further refinement of TD and TO event timings were achieved by tracking the foot in individual camera views. The foot-sized bounding box was projected into each camera view and the region inside each 2D projection was split into vertical slices. A Sobel filter was used to compute the vertical gradient of each slice, and then a horizontal mean reduces each slice to two 1D arrays of colour and gradient values. Frame to frame tracking was performed for each slice by finding the vertical offset that minimises the difference between frames. Tracking begins at the frame where the contact has the largest area in the ground plane occupancy map, typically around mid-stance. TD and TO event timings are determined by tracking from that point forwards or backwards in time to find the frame when the last slice that displaces starts to initiate vertical movement. Ascertaining TD and TO locations and timings permitted the computation of step characteristics (SL, SF, GCT, ST and FT).

Athlete CoM locations, as well as sled CoM locations during pushing, were determined using two separate methods. The first was presented by [11]. This method determines athlete CoM locations by first using a CNN [18] (e.g. DensePose) to segment the torso and head in each 2D image. The CNNs used in these experiments were pre-trained on generic datasets. 2D bounding boxes are created for each head/torso region in each camera view which found to provide an reliable proxy for the true CoM location. An occupancy map is used for initial 3D detection and cross-camera fusion. After initialisation the 3D bounding box for head/torso is optimised for size and location by projecting into the images and optimising the overlap of the projection with 2D segmentations areas. In subsequent frames, the bounding box is initialised in its previous location and re-optimised for the new segmentation, while the occupancy map verifies its continued existence and checks for new detections (S4 Fig in S1 File). Finally, the centroid of each 3D bounding box is passed through a bi-directional Kalman filter to provide an optimal state estimation and the weighted average of both body parts computed. Sled CoM was determined by using a CNN to detect the four corners of the sled (S5 Fig in S1 File) in each 2D camera view. Detections with a high confidence were back projected into the 3D space using the camera calibration and their intersect used to drive the motion of the sled which was treated as a rigid object. Finally, the centroid of the sled was passed through a Kalman filter. Athlete and sled CoM velocities were computed using a finite central differences method before being averaged across each step. The approach proposed by Evans et al. [13] will be referred to in this work as the ‘3D bounding box’ method.

In order to evaluate system performance, the results from the computer vision systems were compared to the criterion system (marker-based motion capture) using linear regression and Bland-Altman analysis showing the mean difference (bias) and 95% limits of agreement (LoA). Further analysis comparing the pushing and sprinting data used estimation statistics [20], which focuses on the magnitude of the effect and its precision. Gardner-Altman estimation plots [21] were produced containing paired Cohen’s d effect sizes and 95% bootstrap confidence intervals (CIs). Five-thousand bootstrap samples were taken and the confidence interval was bias-corrected and accelerated [22]. P values were computed using a Permutation t-test. The reported P values represent the likelihood of observing the effect size, if the null hypothesis of zero difference is true. For each permutation P value, 5000 reshuffles of the control and test variables were performed. Effect sizes and CIs are reported as: effect size [CI width, lower bound; upper bound]. Effects sizes were interpreted according to Cohen’s [23] guidelines (small effect ≥ 0.2, moderate effect ≥ 0.5 and large effect ≥ 0.8).

Results

Motion capture system mean reconstruction accuracy was 0.91 ± 0.76 mm for the criterion system and 0.74 ± 0.68 mm for the computer vision system demonstrating high accuracy for both systems. Validation results for pushing temporal step characteristics are given in Table 1. High levels of agreement were observed for temporal variables between the proposed computer vision system and the criterion system (Table 1), with differences falling within 1.5 frames. Excellent agreement was observed for spatial variables such as SL with mean differences of 0.001 ± 0.012 m.

Table 1

Between system comparison of computed step characteristics during pushing.

Variable	Mean Difference (Bias)	± SD	95% LoA	R2
GCT (s)	0.008	0.015	0.037	0.10
FT (s)	-0.008	0.016	0.023	0.28
ST (s)	0.001	0.017	0.033	0.20
SL (m)	0.001	0.012	0.021	0.99
SF (Hz)	-0.022	0.148	0.278	0.20

For pushing using the 3D bounding box based method, mean differences between the criterion and computer vision based step-averaged sled velocities were -0.015 ± 0.023 m.s-1, while athlete CoM differences were -0.017 ± 0.080 m.s-1 (Table 2). For sprinting using the 3D bounding box method, mean differences for the athlete step-averaged CoM were 0.004 ± 0.068 m.s-1. Bland-Altman and linear regression plots can be found in S6–S13 Figs in S1 File.

Table 2

Comparison of computed CoM step velocities during pushing and sprinting.

Activity	Variable	Mean Difference (Bias)	± SD	95% LoA	R2
Pushing	Athlete CoM Velocity (m.s-1)	-0.017	0.080	0.140	0.94
Pushing	Sled Velocity (m.s-1)	0.015	0.023	0.061	0.99
Sprinting	Athlete CoM Velocity (m.s-1)	0.004	0.068	0.138	1.00

The following results were derived using the experimental computer vision system. Step characteristics as a function of step velocity are given in Fig 2 for both pushing and sprinting.

Fig 2

Step characteristics during sprinting (red) and pushing (blue) shown as a function of step velocity.

Step characteristics inter-limb asymmetries for both pushing and sprinting are given in Figs 3 and 4 as estimation plots showing the paired differences. Results show that when pushing, athletes exhibit similar step characteristics to those observed during sprinting at similar velocity. The results of the estimation statistics including confidence upper and lower bounds and P values are given in S1 Table in S1 File.

Fig 3

Paired step characteristic differences for inside and outside leg during pushing (left column) and the corresponding leg during sprinting (right column).

The paired Cohen’s d is plotted on floating axes on the right as a bootstrap sampling distribution. The Cohen’s d effect size is depicted as a dot; the 95% confidence interval is indicated by the ends of the vertical error bar. Top row—SL, centre row—SF, bottom row—SV.

Fig 4

Paired step characteristic differences for inside and outside leg during pushing (left column) and the corresponding leg during sprinting (right column).

The paired Cohen’s d is plotted on floating axes on the right as a bootstrap sampling distribution. The Cohen’s d effect size is depicted as a dot; the 95% confidence interval is indicated by the ends of the vertical error bar. Top row—GCT, bottom row—FT.

Discussion

The ability to non-invasively capture skeleton push-starts and sprinting kinematics in challenging environments would allow coaches to better understand the key aspects within and between these two training activities. In this study a computer vision and deep learning methodology was developed and validated, with its utility to accurately extract biomechanical characteristics and allow novel, previously-inaccessible information to be captured in a challenging real-word environment demonstrated.

Accuracy of spatial variables such as SL was excellent with a mean difference between systems of 0.001 m. This particular outcome was not surprising as the methods used to calibrate both camera systems and reconstruct 3D locations were fundamentally the same. However, this is of course dependent on each system’s ability to reliably detect and track a location on the foot and demonstrates that this can indeed be achieved in the absence of markers. Temporal variables such as GCT and FT presented higher differences (Table 1) but these were still within approximately 1.5 frames of the criterion system. Due to occlusion caused by the sled, the ground plane with which footfall events were detected was raised by 0.025 m which caused a systematic overestimation of GCT and systematic underestimation of FT as a result of early touch-down detection and late toe-off detection. However, mean differences were similar to those reported for the widely used OptoJump™ system (Microgate, Bolzano, Italy) (0.005 ± 0.004 s, [24]) which also has a raised ground plane. Furthermore, ST (GCT + FT) differences were lower (0.001 ± 0.017 s) as the errors from overestimated GCT and underestimated FT effectively cancel each other out, further indicating that errors in temporal variables were likely due to the elevated ground plane. Alternative approaches using wearable technology such as body or foot mounted accelerometers [25] or in-shoe pressure measurement [26] also offer measurement of temporal but not spatial step characteristics. Mean GCT errors of 0.0017 s have been reported for trunk mounted accelerometers [27] and errors of −0.0067 ± 0.0229 s for in-shoe pressure measurement [28] which fall within the same range of measurement error as the vision based system used in this study. However, both wearable approaches are highly sensitive to changes in walking/running velocity, are untested in the skeleton environment and are ultimately more invasive in nature than a vision-based approach.

Sled velocities and athlete mass centre velocities (Table 2) demonstrated excellent agreement with the marker-based system where mean step-averaged differences of -0.017 m.s-1 for the athlete CoM and -0.015 m.s-1 for the sled were observed. Notably in this work, we demonstrate athlete and sled tracking performance increased upon our previous work [12, 13] with substantial reductions mean difference SDs and substantial increases in R2 values (Table 2). The performance increases were attributed to more robust outlier detection and improved noise process noise predictions in the Kalman filtering stage. Furthermore, the computer vision methods demonstrated superior performance to other field-based methods for measuring athlete running velocities such as laser distance measurement, which exhibits mean errors of up to 0.41 ± 0.18 m.s-1 [29], two to four times higher than the results of this study. Within the application of this study (skeleton push start and sprint accelerations), it appears that the proposed deep learning based computer vision methods provide an alternative, non-invasive way to capture important technique related information (step characteristics and mass centre velocities) where more conventional systems would not be viable. More work is required to reduce measurement errors of some temporal step characteristics (e.g. FT and GCT during pushing) in the proposed markerless methods before they are comparable to marker-based motion capture. However, this field of research is advancing rapidly, and such improvements are likely to emerge in the near future.

Sled pushing step characteristics as a function of distance demonstrated a decrease in SL, SF and FT, and increase in GCT as a result of the external load created by the sled and the bent over postures required to push the sled. These findings align with other studies examining the effects of external resistance (15–30% of body weight) on sprinting performance [28, 30–32] but such studies only considered the overall differences between sprinting with and without external resistance and did not examine how step characteristics between the two activities align as a function of step velocity. Comparison of GCT and FT as a function of step velocity revealed similar results between pushing and sprinting (Fig 2) despite the clear differences in the constraints of these movements. This suggests that sprinting activities at similar velocities to those achieved during pushing broadly mimic these high-level biomechanical variables (step characteristics). Exposure to low resistance sprints training such at sled pushing has been demonstrated to increase the athlete’s ability to develop horizontal forces at higher velocities [33], a training stimulus that the highly trained athletes of this study have been exposed to in large volumes.

The pushing data which were captured between 5 m and 10 m from the starting line, aligned with sprinting data captured between 0 m and 10 m for all athletes suggesting that there may be a greater level of biomechanical specificity earlier in the sprinting phase where postures, velocities and step characteristics more closely match those of pushing. Therefore, it may be more appropriate to match the two activities as a function of velocity and not simply as a function of distance when considering how sprinting activities could be used to replicate pushing. Brazil et al. [34], encouraged coaches to consider the principle of specificity in a holistic way considering all the biomechanical information that is available to ensure that a) the most appropriate training activities are selected and b) the likelihood of achieving positive training adaptation and enhancing performance of the target skill is maximised. Consequently, coaches are now able to make evidence-based decisions to meet the criteria described by Brazil et al. [34], by using the system detailed in this study.

Analysis of the athlete velocities during pushing between approximately 5 m and 15 m and athlete sprinting velocities between 0 m and 10 m further demonstrates that when pushing, the athlete achieves comparable velocities later in the acceleration phase. This is perhaps unsurprising given that during pushing the athlete is highly constrained and must accelerate the sled from a static position. Previously, Colyer et al. [35], has demonstrated very strong associations between push-start performance and both 15–30 m unresisted sprint time and 0–15 m resisted sprint time (sled load, 10 kg for males and 7.5 kg for females). This observation supports the notion of matching sprinting and push activities using CoM velocity and further demonstrates that early acceleration sprints training may be mechanically more specific to pushing, at least for the phase of pushing being analysed in this study.

Comparison of inside leg and outside step characteristics revealed significant asymmetries and large effect sizes during pushing for SL, SF, SV and FT but not for GCT (S1 Table in S1 File). No significant asymmetries were observed for any step characteristic variable during sprinting. The large effect sizes and significant differences observed during pushing are likely the result of the asymmetrical pushing technique where the athlete must push the sled with one arm while the contra-lateral arm is free to swing. Athletes exhibited longer step lengths but lower step frequencies on the inside leg, which typically resulted in higher step velocities. The higher step frequencies on the outside leg were due to reduced flight times as there were no clear differences between GCTs. It is worth noting that the athletes never actually lose full contact with the ground due to the requirement to push and remain in contact with the sled using the support arm. However, athletes appear to spend a similar amount of time imparting an impulse in foot-ground contact, regardless of which leg, inside or outside, is in contact with the ground. During upright sprinting, a lower flight time and step length would indicate lower force production during the preceding contact [36, 37]. The observed asymmetries could, therefore, indicate a compromise in force production capabilities during the outside leg contact. However, this asymmetry could also reflect differences in how effectively energy is transferred to the sled. When generating force on the outside leg, due to the greater distance from the sled to the point of force application on the ground, the transfer of energy to accelerate the sled could conceivably be less efficient than when contacting the ground with the inside leg. Alternatively, it could be that in the bent-over position, athletes are “pulled” (due to the requirement to maintain contact with the sled) towards the ground during flight more on the outside than the inside leg, which would also reduce flight time and step length. A full body kinetic and kinematic analysis of the pushing technique is required to further understand how pushing a sled alters acceleration mechanics and creates the asymmetries observed in this study.

Conclusions

A novel computer vision and deep learning based approach to non-invasively capture kinematic data was thoroughly validated for skeleton push starts. The method was applied in a challenging real-world environment and application (skeleton push starts and sprinting) and was able to capture representative kinematic data including step characteristics and mass centre velocities of the athlete and sled through a fully automated end-to-end workflow. This approach could be employed by coaches and practitioners to monitor technique where traditional motion capture techniques are not practical. The computer vision based system was used to capture skeleton push start and sprinting data where step characteristics were found to be comparable between activities when matched as a function of step velocity. The developed system allowed large asymmetries to be observed between the inside and outside leg during pushing that were not present during sprinting. However, regardless of the leg that is in contact with the ground, inside or outside, the time spent pushing against the ground applying force was not different, potentially indicating compromised outside-leg force production that should be considered by skeleton coaches. These performance observations demonstrate the utility of this system to monitor skeleton athletes’ progress in response to training.

(PDF)

Click here for additional data file.

26 Apr 2021

PONE-D-20-34441

Development, Evaluation and Application of a Novel Markerless Motion Analysis System to Understand Push-Start Technique in Elite Skeleton Athletes

PLOS ONE

Dear Dr. Needham,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Make sure to address all the comments provided by the reviewers, in particular you should clearly highlight your contributions and how they extend those already published in your conference paper [33]. We would also like to remind you that you cannot not reuse text which has already been published.

Please submit your revised manuscript by Jun 10 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Jean-Christophe Nebel, Ph.D

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Thank you for stating in your Funding Statement:

This investigation was part-funded by CAMERA, the Reserch Councils UK Centre for the Analysis of Motion, Entertainment Research and Applications, EP/M023281/1.

Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now.  Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

3. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see http://journals.plos.org/plosone/s/data-availability.

Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories. Any potentially identifying patient information must be fully anonymized.

Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access.

We will update your Data Availability statement to reflect the information you provide in your cover letter.

4. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: http://journals.plos.org/plosone/s/supporting-information.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #3: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: This paper deals with a very interesting application of recent computer vision and deep learning developments. It shows how those development can be applied to real-world problems. It is also very interesting to see the accuracy of those developments in comparison to marker-based methods.

My main concern with this paper is related to the similarity with a previous paper from the same authors:

[33] Needham, L., Evans, M., Cosker, D., & Colyer, S. (2020, March). Using Computer Vision and Deep Learning Methods to Capture Skeleton Push Start Performance. In 38th International Conference on Biomechanics in Sport.

The authors must highlight the differences with their previous work. Some parts of this paper are very similar to [33], particularly Sections "Materials and methods", "Results" and "Discussion". This is most of the paper.

My experience with DensePose and, probably, CPDP (I don't have experience with this model) is that the segmentation is not accurate. The segmentation misses some points in the border of the body. This can be seen in Fig. S1, it seems that the segmented feet are smaller than in the real image. However, this doesn't seem to have an effect in the results presented in this paper. This is probably to the fact that the ground plane is elevated. However, the authors mention that this elevation avoids occlusions with the sled, but I think that this elevation what favours the most is the issue of segmenting smaller feet. In the future this could be solved either with better segmentation methods or by applying a 3D mathematical morphology operation to "recover" those missed points in the border of the body.

- Line 203: It is not clear to me how to read Fig S3 when the authors say "occupancy maps shown on the left side" in the caption.

- Line 210: which colour and gradient based features are computed

Why do you think that the differences between both systems are higher for SF in Table 1?

Line 326: Table S2 is not in the Supplementary material

Line 335: I cannot find a numerical value in the paper or the supplementary material that supports that FT and GCT during pushing are the measures that require a higher reduction of measurement errors.

Line 376: As Table S2 is missing, I don't know if the reference to Table S1 is correct here and, then, I am reading the correct information

Minor comments:

- Line 84: However, in a previous study...

- Line 281: What is SV?

Reviewer #2: This paper presents a study on utilizing some existing computer vision techniques to capture sprinting and skeleton push start step characteristics and mass center velocities.

It is more like a project report rather than a scientific paper. I cannot find its originality at a research paper level. For example, it utilized CDCL-human-part-segmentation [28] and DensePose [21] for detecting and segmenting body parts, and utilized a method in [17] for athlete CoM and sled CoM localization.

It is not suitable for publishing in its present form.

Reviewer #3: In this paper, the authors present and evaluate a deep-learning-based stereovision system in the context of elite skeleton. The obtained results are very promising for step length estimation and velocities. However, there are room for improvement for temporal variables.

The authors also present first motion analysis results which i found a bit confusing since the accuracy of the system for such dynamic action still needs to be established and some points of the implementation of the system are superficially described.

My specific comments are below.

1. The pixel resolution of captured images from the computer vision system is not specified in the text.

2. Fig 1 should be made clearer with some more annotations, for instance about starting line and direction of the athlete.

3. Some more details about calibration would be needed. What was the calibration pattern for extrinsic parameters computation ? What was the diameter of the marker for alignment of systems' coordinate systems ? What was the tracking technique for the videos ? Same question for the wand.

4. Lines 193 - 199 contain elements of discussion. The authors justify the use of a DL approach but this should be discussed later in the paper.

5. Line 199 "To detect approximate foot contact locations and timings, foreground segmentations were fused and occupancy maps of the ground plane were computed". This sentence seems to me essential to assess the originality of the approach, but the interested reader has not enough elements to fully understand it. I think the authors should develop a bit on this point.

6. The authors mentioned that the sled corners were detected thanks to a CNN. Which one ? How it was trained ?

7. Similar question about the network used for COM inference. On which database has it been trained ? I think the authors should discuss the need (or not) of collecting a specific database for such dynamic motion.

8. The triangulation method is superficially described and if I'm not mistaken the 3D inference of COM was not part of ref 28.

9. What is the accuracy of COM localization when compared to qualisys ?

10. The authors claim in the abstract that "temporal variable were within 1.5 frames of the criterion measures". This statement is wrong since it is a mean (not-signed) difference between the two system. The uncertainty is between 0.02 and 0.04. So the difference between the two systems for GCT can be 8 frames. The authors should discuss these errors in the context of an elite practice.

11. An important aspect of any motion analysis system is not treated or discussed : inter-trial variability of the computed parameters.

12. Minor comment, in table 1. The unit for SF should be Hz.

13. While the first part of the results section is well balanced between the tables and the text (good synthesis of the tables in the text), FiG2 to 4 are not synthetized at all in the text.

14. It is interesting that the authors develop on some findings on their trial and new system. But since it is a novel system, it would be appreciable to see if these results could be verified on the ground-truth system.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

11 Jun 2021

Response to Reviewer #1:

Thank you for your helpful comments. We have replied below in red.

Reviewer #1: This paper deals with a very interesting application of recent computer vision and deep learning developments. It shows how those development can be applied to real-world problems. It is also very interesting to see the accuracy of those developments in comparison to marker-based methods.

My main concern with this paper is related to the similarity with a previous paper from the same authors:

[33] Needham, L., Evans, M., Cosker, D., & Colyer, S. (2020, March). Using Computer Vision and Deep Learning Methods to Capture Skeleton Push Start Performance. In 38th International Conference on Biomechanics in Sport.

The authors must highlight the differences with their previous work. Some parts of this paper are very similar to [33], particularly Sections "Materials and methods", "Results" and "Discussion". This is most of the paper.

The conference abstract in question (presented at International Society of Biomechanics in Sports annual conference) was a pilot study and represented an early version of our vison-based system. It should also be noted that this work was not peer reviewed to the same level that a paper might be in a journal. In the current work we present a later iteration of our methods which have evolved substantially in terms of how the system works and its performance. To address these concerns, we have added the following information:

Additional technical details have been added to the methods section for clarity. Lines 200-250

In the results section you will see some important performance increases when compared to our pilot work. For instance, random error between our proposed system and criterion system for athlete CoM velocity and sled CoM velocity were reduced from 0.186 m.s−1 and 0.133 m.s−1 to 0.08 m.s−1 and 0.023 m.s−1 respectively. Such performance improvements substantially improved our systems ability to detect small, meaningful changes in push start and sprinting performance. Additionally, our validation in the present manuscript is substantially more detailed with more trials and multiple exercise modalities being included (pushing and sprinting). Furthermore, we provide full Bland-Altman analyses and linear fit models for all variables assessed, allowing the reader to make a more informed decision regarding the statistical evaluation of our work. This paper is multidisciplinary in nature and provides extensive evaluation against a gold standard motion capture method in biomechanics research. This was an essential step to demonstrate to the biomechanics research community, in particular, that markerless techniques can perform to comparable levels with established marker-based techniques.

This in turn allows us to provide real-world application of our approach to provide novel biomechanical insight to inform Olympic coaches and athletes about skeleton pushing technique which is discussed in detail in the discussion.

My experience with DensePose and, probably, CPDP (I don't have experience with this model) is that the segmentation is not accurate. The segmentation misses some points in the border of the body. This can be seen in Fig. S1, it seems that the segmented feet are smaller than in the real image. However, this doesn't seem to have an effect in the results presented in this paper. This is probably to the fact that the ground plane is elevated. However, the authors mention that this elevation avoids occlusions with the sled, but I think that this elevation what favours the most is the issue of segmenting smaller feet. In the future this could be solved either with better segmentation methods or by applying a 3D mathematical morphology operation to "recover" those missed points in the border of the body.

We too observe that these segmentation algorithms can produce segmentations that are smaller than the actual area of the person or incomplete, however, the exact area of the segmentation does not directly matter. In the occupancy map stage, the most important thing is that there is sufficient segmentation information to identify the presence of the foot (one might say we need an accurate segmentation, but not an especially precise/detailed one). For localisation the segmentation must give enough of an indication of the volume/position of the foot to guide the optimisation of the 3D bounding box location, and there are various elements in the design of the optimisation which mitigate partial segmentations. Obviously, there are limits, and as we develop our system we are considering ways of getting better segmentations from these and other algorithms - but those developments are for future work and not part of the presented work. It should also be pointed out that the elevated plane does not affect the optimisation of the 3D bounding box that provides the precise step localisation – that still uses a bounding box assumed to be on the z=0 plane.

- Line 203: It is not clear to me how to read Fig S3 when the authors say "occupancy maps shown on the left side" in the caption.

We have re-worded the caption and additionally provided further technical detail regarding how occupancy maps are used in the methods section. Lines 200-207.

- Line 210: which colour and gradient based features are computed

A Sobel filter was used to compute the vertical gradient of each slice, and then a horizontal mean reduces each slice to two 1D arrays of colour and gradient values. Lines 213-215.

Why do you think that the differences between both systems are higher for SF in Table 1?

Step Frequency (SF) refers to the cadence or number of steps per second. In this study we computed SF as 1/ST where ST (step time) was computed as the sum the GCT and FT. As such for both marker-based and markerless methods, SF is highly sensitive to any changes in GCT or FT measurements. We believe that this high sensitivity to differences or changes in one or both measurement system may have contributed to the higher differences reported.

Line 326: Table S2 is not in the Supplementary material

This should in fact refer to Table 2 in the results section. Typo corrected.

Line 335: I cannot find a numerical value in the paper or the supplementary material that supports that FT and GCT during pushing are the measures that require a higher reduction of measurement errors.

Our mean differences (bias) between for temporal variables such as GCT and FT were low (Table 1), however, the LoA demonstrated that the range of likely values from future measurements is still higher than those seen for marker-based methods. This notion is further supported by the lower R2 values.

Line 376: As Table S2 is missing, I don't know if the reference to Table S1 is correct here and, then, I am reading the correct information

Yes, this is referring to the correct table. Additional detail has been added to the table caption.

Minor comments:

- Line 84: However, in a previous study...

Sentence restructured for clarity.

- Line 281: What is SV?

Step velocity (SV) – definition added to methods section line 189.

Response to Reviewer #2:

Thank you for your comments. We have replied below in red.

Reviewer #2: This paper presents a study on utilizing some existing computer vision techniques to capture sprinting and skeleton push start step characteristics and mass center velocities.

It is more like a project report rather than a scientific paper. I cannot find its originality at a research paper level. For example, it utilized CDCL-human-part-segmentation [28] and DensePose [21] for detecting and segmenting body parts, and utilized a method in [17] for athlete CoM and sled CoM localization.

The present manuscript is not simply a vision based paper but rather provides a multidisciplinary approach utilising knowledge from computer vision, deep learning and sports biomechanics. We present the latest iteration of our system for non-invasively measuring sprinting and pushing technique before rigorously validating this method against the de-facto gold standard and finally applying the method to solve a real-world problem. The method presented allowed us to obtain novel results of the first biomechanics analysis of skeleton push starts and compare directly to sprinting technique in an elite population. Given all of the above, we respectfully disagree and believe that the originality of this scientific study is clear.

It is not suitable for publishing in its present form.

Response to Reviewer #3:

Thank you for your helpful comments. We have replied below in red.

Reviewer #3: In this paper, the authors present and evaluate a deep-learning-based stereovision system in the context of elite skeleton. The obtained results are very promising for step length estimation and velocities. However, there are room for improvement for temporal variables.

The authors also present first motion analysis results which i found a bit confusing since the accuracy of the system for such dynamic action still needs to be established and some points of the implementation of the system are superficially described.

If we understand your comment correctly, you have concerns over the accuracy of the marker-based system in these types of movements. Whilst we acknowledge that there is a small degree of measurement error in the marker-based system, most notably due to skin movement artifact, such limitations are well studied and can be effectively mitigated. As such marker-based motion capture systems remain the gold standard in this field and thus provide s us with a clear criterion to evaluate against. Further technical detail pertaining to the implementation of the system have been added to the methods section. Lines 165-239.

My specific comments are below.

1. The pixel resolution of captured images from the computer vision system is not specified in the text.

The JAI machine vision cameras record with an image resolution of 1920 × 1080 pixels – This has been added to the methods section. Lines 144-145.

2. Fig 1 should be made clearer with some more annotations, for instance about starting line and direction of the athlete.

Figure and caption updated to provide more information. See methods section, lines 159-162.

3. Some more details about calibration would be needed. What was the calibration pattern for extrinsic parameters computation ? What was the diameter of the marker for alignment of systems' coordinate systems ? What was the tracking technique for the videos ? Same question for the wand.

Calibration of the machine vision cameras used a calibration board with 10 x 9 78.5 mm black circles on it. Observations of the calibration board were used to initialise each camera’s intrinsic parameters (Zhang, 2000) before extrinsic parameters were initialised from pairs of cameras with shared board observations. A global optimisation was performed using Sparse Bundle Adjustment (Triggs et al., 2003) to determine the final intrinsic and extrinsic parameters. Alignment of the two camera systems coordinate systems tracked a 16 mm retro-reflective motion capture marker which was tracked manually in the machine vision system’s image data. Further detail has been added to the methods section.

4. Lines 193 - 199 contain elements of discussion. The authors justify the use of a DL approach but this should be discussed later in the paper.

Discussion of segmentation and background subtraction techniques is outside the scope of this paper. The information provided in the methods is designed a) distinguish our method from previous approaches and b) provide the reader with some insight into our thought process during development. Therefore, we believe these do warrant the further explanation in the methods section.

5. Line 199 "To detect approximate foot contact locations and timings, foreground segmentations were fused and occupancy maps of the ground plane were computed". This sentence seems to me essential to assess the originality of the approach, but the interested reader has not enough elements to fully understand it. I think the authors should develop a bit on this point.¬¬

Further detail has been added to the methods section. Lines 199-220.

6. The authors mentioned that the sled corners were detected thanks to a CNN. Which one ? How it was trained ?

To detect the corners of the sled a keypoint detection model based on DeeperCut with a ResNet-150 backbone, was trained via transfer learning using approximately 200 annotated images. This was achieved using a modified version of the DeepLabCut framework. Further detail has been added to the methods section. Line 222-239.

7. Similar question about the network used for COM inference. On which database has it been trained ? I think the authors should discuss the need (or not) of collecting a specific database for such dynamic motion.

The segmentation networks have been taken “off the shelf”, and have been either CDCL or DensePose – they are not trained on skeleton specific datasets but have generally performed adequately, but one would clearly expect that training on skeleton and sprinting specific data would offer some advantage in segmentation quality, but how much that would improve tracking quality is less obvious. This is explained in the lines 194-200.

8. The triangulation method is superficially described and if I'm not mistaken the 3D inference of COM was not part of ref 28.

We have clarified the text of our description regarding the 3D reconstruction approach, which is fully described in ref: Line199-201 & 234 – 236.

9. What is the accuracy of COM localization when compared to qualisys ?

The fused bounding box method is not designed to provide a measure of CoM position that is directly comparable with the marker-based CoM position. The variable of interest to end users (e.g. biomechanists, coaches and athletes) is the CoM velocity. As such our method uses a weighted average of the fused torso and head bounding box centroids whose derivative provides an excellent proxy for the true CoM velocity (e.g. Table 2). Further detail regarding this has been added to the methods section (Lines 230-251).

10. The authors claim in the abstract that "temporal variable were within 1.5 frames of the criterion measures". This statement is wrong since it is a mean (not-signed) difference between the two system. The uncertainty is between 0.02 and 0.04. So the difference between the two systems for GCT can be 8 frames. The authors should discuss these errors in the context of an elite practice.

We have reworded this sentence to acknowledge that on average, values were generally within ± 1.5 frames of the ground truth measure. This value was derived from the Bland-Altman analysis bias and as such is in fact a signed mean. Results, including the bias and random error, are discussed in relation to the performance of other measurement technologies to which we demonstrate comparable performance. It should also be noted that while such technologies are commonly used to collect running based data they cannot be used in the challenging skeleton training environments. This is discussed in lines 325-344.

11. An important aspect of any motion analysis system is not treated or discussed : inter-trial variability of the computed parameters.

Inter-trial variability of the differences between the measurement systems is presented in the Bland-Altman analyses. I.e. the SD of bias and the 95% LoA. Biological variability of the participants is outside the scope of this study but could certainly be assessed in future studies and/or taken into account when evaluating longitudinal changes in these parameters.

12. Minor comment, in table 1. The unit for SF should be Hz.

Thank you, typo corrected.

13. While the first part of the results section is well balanced between the tables and the text (good synthesis of the tables in the text), FiG2 to 4 are not synthetized at all in the text.

Further detail has been added to the results section. Lines 282 – 283.

14. It is interesting that the authors develop on some findings on their trial and new system. But since it is a novel system, it would be appreciable to see if these results could be verified on the ground-truth system.

The purpose of this paper was to present a detailed and robust evaluation of our novel system before discussing the biomechanics information provided and how it relates to both sprinting and pushing techniques. The results presented in Figs 2 – 4 are the same results that have been used to verify our novel system against the ground-truth system in Table 1 -2, Table S1 and Figs S6 – 13. Thus if we have understood your comment correctly, this verification has already been performed.

11 Jul 2021

PONE-D-20-34441R1

Development, Evaluation and Application of a Novel Markerless Motion Analysis System to Understand Push-Start Technique in Elite Skeleton Athletes

PLOS ONE

Dear Dr. Needham,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

If you could address the remaining minor comments, we would be happy to publish your manuscript.

Please submit your revised manuscript by Aug 25 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Jean-Christophe Nebel, Ph.D

Academic Editor

PLOS ONE

Journal Requirements:

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #2: (No Response)

Reviewer #3: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #2: Yes

Reviewer #3: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #2: Yes

Reviewer #3: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #2: Yes

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #2: Yes

Reviewer #3: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #2: I have read the revised paper and the response to the reviewers. It is a good project report but not a scientific research paper.

Reviewer #3: The authors have well adressed all my comments except 2, that remained not understood by the authors but also maybe because of the lack of clarity of these 2 comments.

The 2 comments are actually linked.

Comment 1 : "The authors also present first motion analysis results which i found a bit confusing

since the accuracy of the system {i precise today i was talking about the markerless system] for such dynamic action still needs to be established"

Comment 2 : "It is interesting that the authors develop on some findings on their trial and new

system. But since it is a novel system, it would be appreciable to see if these results

could be verified on the ground-truth system."

My concern with these 2 comments is that biomechanical interpretations are given thanks to the results of the markerless system. While the validity of the system has been quantified previously against a marker-based system, would the conclusions drawn by the authors would have been exactly the same if the data coming from the marker-based system were used instead ? The assymetries that are depicted are that clear that i presume it is the case but if so, the authors could mention it.

I have a last question. Why comparison results for GCT, FT etc are given only for pushing and not on sprinting ? These variables are analyzed in fig 3 and 4 but we do not know their accuracy.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #2: No

Reviewer #3: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

14 Jul 2021

Reviewer #2:

I have read the revised paper and the response to the reviewers. It is a good project report but not a scientific research paper.

We respectfully disagree with this opinion. In addition to the novel scientific contributions which we outlined in our previous response, the paper in question demonstrates that computer vision based technologies can be applied to have real-world impact. This potential impact could fall across a range of science and medicine disciplines, beyond the specific application (or ‘project’) of skeleton studied here. For example, the principles demonstrated in our paper could be applied by motor rehabilitation scientists and clinicians who quantify movement to inform rehabilitation design and evaluate the effects of disease and treatments. Neuroscientists studying brain-movement interaction and motor learning. Similarly, psychologists could examine human motor development, human motor behavior and the effects of psychological disorders on movement. Sports and exercise physiologists and biomechanists could examine the metabolic costs of human movement, sports techniques (as in our paper), injury mechanisms and equipment design. Finally, engineers could quantify movement for prosthetics, exoskeleton, and rehabilitation robotics design. The scope and breadth of these examples demonstrate the substantial impact that using computer vision technologies applied to human movement research can makes to science and medicine. As such, we feel that we have provided a clear and justified rationale for the scientific contribution this paper makes and why it is suitable for publication in PLOS One.

Reviewer #3:

The authors have well addressed all my comments except 2, that remained not understood by the authors but also maybe because of the lack of clarity of these 2 comments.

The 2 comments are actually linked.

Comment 1: "The authors also present first motion analysis results which i found a bit confusing since the accuracy of the system {i precise today i was talking about the markerless system] for such dynamic action still needs to be established"

Comment 2: "It is interesting that the authors develop on some findings on their trial and new system. But since it is a novel system, it would be appreciable to see if these results could be verified on the ground-truth system."

My concern with these 2 comments is that biomechanical interpretations are given thanks to the results of the markerless system. While the validity of the system has been quantified previously against a marker-based system, would the conclusions drawn by the authors would have been exactly the same if the data coming from the marker-based system were used instead? The assymetries that are depicted are that clear that i presume it is the case but if so, the authors could mention it.

Thank you for clarifying these specific comments. For clarification, any variable that has been presented in this study using our markerless system, has been verified against the ground-truth data (marker-based motion capture). Any systematic or random differences that were reported between the two measurement systems were shown to be small enough to allow for meaningful changes in push start and sprinting technique to be detected. Differences between the systems and thus the accuracy of our markerless system, in relation to the ground truth (marker-based motion capture) are presented in Tables 1 and 2 as well as in the Supporting Information.

I have a last question. Why comparison results for GCT, FT etc are given only for pushing and not on sprinting? These variables are analyzed in fig 3 and 4 but we do not know their accuracy.

As is discussed in the methods section, sprinting validation results for GCT, FT etc are presented in our groups previous work using an earlier iteration of the methods presented in this study. For full details see Evans M, Colyer S, Cosker D, Salo A, Ieee, editors. Foot Contact Timings and Step Length for Sprint Training. 18th IEEE Winter Conference on Applications of Computer Vision (WACV); 2018 Mar 12-15; Nv2018.

Submitted filename: Response_to_Reviewers_rev2.pdf

Click here for additional data file.

25 Oct 2021

Development, Evaluation and Application of a Novel Markerless Motion Analysis System to Understand Push-Start Technique in Elite Skeleton Athletes

PONE-D-20-34441R2

Dear Dr. Needham,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Leonardo A. Peyré-Tartaruga, Ph.D.

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Congrats Dr. Laurie, although the reviewers have raised some points, you have been done a good job here, and the paper is ready for publishing.

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #3: All comments have been addressed

Reviewer #4: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #3: Yes

Reviewer #4: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #3: Yes

Reviewer #4: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #3: (No Response)

Reviewer #4: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #3: Yes

Reviewer #4: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #3: All comments have been adressed. I recommend acceptance of this paper for publication in plos one

Reviewer #4: The authors had two aims: 1) Develop and compare a markerless motion analysis system (deep learning) with a traditional marker-based motion system performances during the tasks of skeleton push-track and sprinting. 2) Compare biomechanical variables of skeleton puh-track vs. sprinting by healthy adults using the new markerless motion analysis system.

The markerless method development is well explained, but some questions appear at Results and Discussion section.

As major point, the three paragraphs from discussion (Lines 367 – 401) discuss data not presented in the Results section. Despite the relevance from the arguments brought, I think that the Discussion thinking line should be supported by the data presented previously. Below the details of each paragraph is depict, but maybe the authors should review these arguments in overall.

The paragraph from Line 367 refers to the variation of spatiotemporal parameters of sled pushing in function of distance, but the data and figure 2 only present the variation of these parameters in function of step velocity.

The paragraphs from Lines 381 and 393 brings informations about to spatial segments from the pushing and sprinting trials, despite these specific data are not presented in Results section.

Perhaps if the authors could exhibit these informations in more details at Results (perhaps Supplementary Material), the reader’s capacity to follow the arguments would be improved.

Minor points

Line 106 – the IMU acronym definition seem unnecessary, considering that it does not appear again in the document.

Line 186 – same as above for the IK acronym.

Line 187 – reference 15 is not “de Leva”. Actually, “de Leva” is not in reference list.

Line 192 – I am in doubt if the authors analyzed a step or stride cycle from the gait: considering the step as the cycle between the same event repetition from ipsilateral to contralateral leg, while stride cycle as the cycle between the same event repetition from ipsilateral-to-ipsilateral leg (Whittle, 2007). Because if the authors refer to step cycle, it seems a bit odd to have both ground contact time and flight time from the same leg measured.

Line 201 – CNN acronym was already defined in Introduction section line 110.

Line 234 – Maybe it would increase the readability to write the authors’ name before reference 11 number, as made in Line 197.

Line 274 – It is not clear what was the authors’ criteria to consider the agreement as “excellent”.

Table 1 and 2 – I could not find the tables’ footnotes to make them self-supporting, explaining the abbreviations, per example.

Line 280 and 282 – it seems to lack the word “speed” after CoM.

Line 297 – I assume that the sentence “Results show that when (…)” refers to Figure 2. But the way that is presented - with the preceding sentence calling Figures 3 and 4 - make a misunderstanding to where to check visually the information. Perhaps explicitly calling Figure 2 in this sentence would fix this.

Figure 3 and 4 – the Coehn’s horizontal lines of zero and mean values are overlapping the left part from the figures; I think that this do not has any valid meaning. Maybe erasing this excess of line would help the figure to come clearer.

Line 325 – CV acronym was not defined before.

Line 339 – considering that ST = GCT + FT, I think that the gait cycle analyzed was stride.

Line 361 - “mass center” means CoM? Anyway, “center” is misspelling.

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #3: No

Reviewer #4: Yes: André Ivaniski-Mello

5 Nov 2021

PONE-D-20-34441R2

Development, Evaluation and Application of a Novel Markerless Motion Analysis System to Understand Push-Start Technique in Elite Skeleton Athletes

Dear Dr. Needham:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Professor Leonardo A. Peyré-Tartaruga

Academic Editor

PLOS ONE