2Trax3: Raising Accessibility and Everyday Use of Automatic Motion
Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation
Algorithms
Samir Duvelek
1
, Dominik Hoelbling
1 a
, René Baranyi
1 b
, Roland Breiteneder
1
, Karl Pinter
1 c
and Thomas Grechenig
1,2
1
Research Group for Industrial Software (INSO), Vienna University of Technology, Vienna, Austria
2
RISE Institute of Technology, Sri Sathya Sai District, Andhra Pradesh, India
Keywords:
Motion Capturing, Video Analysis, Kinematic, Martial Arts, Kicking Techniques, Artificial Intelligence.
Abstract:
A sound technique forms the fundamental basis for many sports, particularly Martial Arts, as it often distin-
guishes between successful hits and being hit. However, the process of improving one’s technique is highly
intricate, often requiring expert feedback and expensive technology such as 3D motion capturing. The integra-
tion of automated technique analysis has the potential to streamline this process and make it more accessible.
In this study, the aim is to democratize technique analysis by developing and evaluating a web application.
This application allows users to upload 2D video recordings of themselves performing the double side kick
technique and receive immediate feedback. To validate the analysis generated by the application, it was com-
pared to a Vicon motion app 3D analysis of the same data from a preliminary study involving 44 participants.
The results of Bland-Altman plot analysis demonstrated a highly significant agreement between the 3D and
2D performance indicators (Mean differences: relative phase duration: <0.04s; vector spreading angle: <15
degrees; relative body position <13%), indicating that the web application is a suitable tool for fast and effec-
tive motion analysis.
1 INTRODUCTION
High performance in sports particularly requires ex-
tensive technique skills among other abilities. Es-
pecially, in combat sports, technique execution does
not solely serve to reach a biomechanical goal but
rather has a tactical purpose (Hoelbling et al., 2021b).
Studying your opponents and understanding their
strengths and weaknesses in advance is also essen-
tial. Once the opponent’s technique has been analysed
correctly, a counter-strategy can be applied (Ouergui
et al., 2021).
Furthermore it can also be used for self improve-
ment, a technique analysis enables targeted training
planning. Certain weaknesses, such as the execution
of certain kicks, can be eliminated through analysis
and subsequent targeted training.
The analysis of such sports techniques has been
the focus of biomechanical and kinematic research
a
https://orcid.org/0000-0001-7099-2576
b
https://orcid.org/0000-0002-0088-9140
c
https://orcid.org/0000-0002-8930-1875
for decades (Elliott, 1999). Even AI applications
continually gain popularity in the field (Lapham and
Bartlett, 1995).
Technique analysis in combat sports is based on
a precise study of techniques, such as punches and
kicks. Each technique requires accurate execution of
sequential actions in combination with correct posture
and precise targeting to achieve maximum effective-
ness (Ambro
˙
zy et al., 2020). Careful analysis and po-
tential adaption can be greatly improve the skills and
abilities of athletes by uncovering physical and coor-
dinative weak spots, which can be compensated in the
training process.
However, despite the multitude of advantages of
technique analysis, it requires highly professionalized
performers and is a very time-consuming task. Partic-
ularly, when it comes to complex motions, which are
rapidly executed with quite little deviations leading to
the distinction between success and failure in fight-
ing situations, a complex multidimensional approach
is necessary to comprehensively disclose the optimum
and define the necessary adaptions (Lees, 2002).
128
Duvelek, S., Hoelbling, D., Baranyi, R., Breiteneder, R., Pinter, K. and Grechenig, T.
2Trax3: Raising Accessibility and Everyday Use of Automatic Motion Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation Algorithms.
DOI: 10.5220/0012165200003587
In Proceedings of the 11th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2023), pages 128-135
ISBN: 978-989-758-673-6; ISSN: 2184-3201
Copyright © 2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
1.1 Double Side Kick
A number of current publications of the double side
kick in pointfighting kickboxing (Yuncza, 2011) for
example shows how complex this process really is
and how many factors must be taken into account to
model high performance (Hölbling et al., 2017). As
a gold standard of this type of analysis researchers
use i.e. Vicon 3D motion capturing systems (Barris
and Button, 2008). Late research successfully de-
fines a number of characteristics, strongly correlat-
ing with performance, measured by both, expert rat-
ings and general competition success (Hölbling et al.,
2020a) (Hoelbling et al., 2021b). However, extract-
ing these variables is practically impossible for regu-
lar coaches, as it requires a very high skill set, very
expensive 3D motion capturing technology and a suf-
ficient amount of time. In addition, cameras can pick
up more detailed characteristics than visible to the hu-
man eye (Barris and Button, 2008).
Some of the study’s exploring the double side kick
used 3D cameras to record various athletes perform-
ing the technique. The technique was then semi-
automatically analysed using the 3D video recordings
(Hölbling et al., 2020b). The performances were si-
multaneously recorded using 2D cameras.
1.2 Preexisting Dataset
A comprehensive data set from a preliminary study
was provided as the foundation for this software ap-
plication (Hölbling et al., 2017). The experiment
recordings consisted of 3D Vicon motion capturing
data with corresponding 2D videos of the double side
kick executions of 44 athletes. All results of the 3D
motion capturing, were extracted and published by
previous researchers (Hölbling et al., 2017) (Hölbling
et al., 2020a) (Hölbling et al., 2020a).
1.3 Automated Analysis Using 2D Video
Given the opportunity of existing scientifically proven
performance characteristics in combination with cur-
rent advancement in the field of AI expert systems
(Zhang and Lu, 2021) and user-based development
(Zorzetti et al., 2022) it appears possible to simplify
this process in a way that regular coaches and athletes
can easily track their performance status and progress.
In particular, these components could be used to de-
velop a system to automatically analyse such tech-
niques with known performance characteristics using
2D videos, advanced motion capturing algorithms and
a set of additional rules to significantly decrease anal-
ysis time effort and sophisticated equipment. There
are already studies that compare the use of 3D and
2D motion capturing in other sports, such as cross-
country running (Maykut et al., 2015) and youth base-
ball (DeFroda et al., 2021), but no research for auto-
matic or semi-automatic analysis of Martial Arts tech-
niques was found.
1.4 Aims
For the purpose of simplifying complex analysis and
ultimately allow for integration of these methods into
regular training, the aim of this study is to develop and
evaluate a system for automatic analysis of 2D videos
of the double side kick, based on performance char-
acteristics defined in (Hölbling et al., 2017) (Hölbling
et al., 2020a) (Hölbling et al., 2020b).
1.5 Hypothesis
Based on existing raw data from the fundamental pub-
lications, following hypothesis can be defined. There
is a significant agreement between the following per-
formance characteristics extracted in 2D (by the sys-
tem) and 3D (by the fundamental research) kinematic
analysis: (a) the duration of a specific segment of the
technique execution relative to the total execution du-
ration, (b) the vertical knee height at specific points
of the technique execution relative to the height of
the trochanter major in neutral standing position, (c)
the distance between knee and frontal shoulder at spe-
cific points of the technique execution relative to the
same distance in neutral standing position, (d) the an-
gle created by the hip and knee of the standing leg
with the vector connecting the knee and hip of the
kicking leg at specific points of the technique execu-
tion, (e) the angle created by the hip and heel of the
standing leg with the vector connecting the heel and
hip of the kicking leg at specific points of the tech-
nique execution, (f) the velocity of the vertical knee
elevation over the course of a specific segment of the
technique execution, (g) the velocity of the kick leg
over the course of a specific segment of the technique
execution.
2 METHODS
Requirements Engineering: In a first step after liter-
ature research and qualitative dataset analysis, semi-
structured interviews (Adams, 2015) with domain ex-
perts in pointfighting kickboxing and biomechanics
were conducted and analysed via qualitative content
analyses (Mayring et al., 2004), to extract essential
requirements.
2Trax3: Raising Accessibility and Everyday Use of Automatic Motion Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation
Algorithms
129
Table 1: Node and phases of the double side kick.
Nodes
Description
Node 1 - Initialisation (INI)
Defined as the moment when the
kicking legs’ foot loses contact with the ground
Phase 1 - Chambering 1 (CH1)
Consists of flexion of the knee as well as flexion and abduction
of the hip joint of the kicking leg
Node 2 - Time of measurement 1 (MT1)
Defined as the highest elevation
of the knee before the kicking legs’ knee angle surpasses 110°
Phase 2 - Kicking phase 1 (KI1)
Mainly consists of extension of
knee and hip joint with ankle flexion of the kick leg
Node 3 - Knee extension maximum 1 (KE1)
The first kick ends with the maximum
extension of the kick leg’s knee
Phase 3 - Chambering 2 (CH2)
Re-Chambering is similarly defined as CH1
and describes the preparation for the second kick
Node 4 - Time of measurement 2 (MT2)
Similarily defined as MT1,
but after the first kick
Phase 4 - Kicking phase 2 (KI2)
Second kicking Phase,
with the same definition as KI1
Node 5 - Knee extension maximum 2 (KE2)
Maximum extension of the
kicking legs’ knee and hip, leading to target contact
These requirements were later adapted and
extended based on observations and comments
throughout the development.
System Design and Implementation: Based on the
extracted requirements a frontend and backend archi-
tecture was designed and an application was devel-
oped using the method of prototyping (Floyd, 1984).
This procedure allowed practical demonstration of
relevant parts of the system early on in the develop-
ment process. Multiple iterations, incorporating ex-
pert feedback and trial and error processes were per-
formed, to improve user experience and measurement
accuracy.
Phase and Node Definition: The foundation for
motion analysis is the segmentation (Meinel and
Schnabel, 2007; Göhner, 1992) of the double side
kick into six functional phases and seven nodes as
defined by Hölbling et al. (Hölbling et al., 2017).
Nodes are defined as the exact moments when one
phase transitions into another, see table Table 1.
Variable Definition: After the definition and
segmentation of the movement, the performance
indicators had to be extracted and calculated at the
nodes or within the phases, as described in (Hölbling
et al., 2017). They are grouped into the following
three categories: (a) the relative duration of the
functional phases, (Hölbling et al., 2017) (b) the
relative position of relevant body parts (kicking legs’
knee height and distance between knee and front
shoulder) , (c) the accumulated angles between the
legs at the time of a node, (Hölbling et al., 2020a)
and (d) the velocity of a body parts motion during
a phase (Hölbling et al., 2020b), see table Table 2.
The time-domain parameter values in the following
text are the absolute duration given in seconds.
The angles are given in degrees. And the distance
parameter values are given as relative values without
unit.
Phase Detection and Motion Analysis Algorithms:
Based on the phase, node and performance indicator
definitions, an advanced algorithm was developed,
which solely relied on some anthropometric data
and was able to automatically separate the phases
and extract the variables. In an iterative process, the
algorithm was then improved.
Statistical Evaluation: In a final step, the extracted
data (from the 2D analysis) was statistically compared
to the pre-existing data (from Vicon 3D motion cap-
turing). Statistical analysis and plot generation was
conducted using the Python programming language
Python Software Foundation. Python Language Ref-
erence, version 3.9).The differences between 3D and
2D variables were checked for normal distribution by
using Shapiro-Wilk tests. Similar to previous stud-
ies exploring the agreement between 3D and 2D mea-
surements in kinematic analysis (Peebles et al., 2021;
Schurr et al., 2017), Bland-Altman plots (Bland and
Altman, 1986) were calculated for each of the de-
pendent variables, in order to evaluate agreement be-
tween 3D and 2D analysis.
The y axis is constructed using the average mean
difference, because the x displays a small concentra-
icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support
130
Table 2: Performance indicators.
Durations
Relative duration of each phase,
normalized by the duration from the nodes INI to KE2
Distances
(i) The vertical height of the kick leg’s knee normalized by
trochanter major height in straight stance (KHK).
(ii) The relative distance between the kick leg’s knee and nearest shoulder normalized
by their distance in neutral standing position. Both extracted at every node
Vector spreading Angle (VSA)
(i) The angle between the femur bones of
both legs (vector connecting knee and hips’ center of rotation).
(ii) The angle between both legs (vector connecting hip and ankles’ center of rotation)
Velocity
The mean velocity of the kick leg’s knee vertical elevation during CH1 and CH2.
And the mean velocity of the kick legs’ foot in target direction during phase KI2
tion range (Bland and Altman, 1999). For 95% limits
of agreement were used.
For the angle characteristics the a priori accept-
able limits were set to 20
. This decision was made
by consulting the results of similar studies where a
subset of the results exceeds 20
(Schurr et al., 2017)
or where an average of 20
agreement (Remedios and
Fischer, 2021) was reported.
For the duration and distance characteristics there
were no comparable studies found, therefore it has
been decided to calculate an equivalent of the 20
limits of agreement taken from (Remedios and Fis-
cher, 2021). It was decided to consider the minimum
and maximum value for each of the individual charac-
teristics and choose the 25 % value of the difference
between minimum and maximum as the acceptable
limit for that characteristic. This choice was made to
ensure that in the case that, the 3D analysis measures
a performer to be either in the top 25 % or bottom 25
% of a characteristic, the 2D analysis agrees with the
3D analysis in so far that it puts the performer into the
top half or bottom half for that characteristic respec-
tively. This is defined as the minimum viable case for
a characteristic to be considered in the 2D analysis.
This means for the duration variables the limits
are for CH1 (0.162s), KI1 (0.05s), CH2 (0.19s), KI2
(0.108s).
And for the distance variables: For KHK1 (22),
KI1 (28.5), KHK2 (16.5), DKS1 (15.5), DKS2 (20)
3 RESULTS
The gathered results will be described in the following
subsections.
3.1 Requirements List
The iterative requirements engineering process re-
sulted in a set of requirements, which are categorized
into the following three groups: (a) User instruction.
The user shall be instructed in how to execute the
technique and how to film the technique being per-
formed. (b) User input. The user shall be enabled
to upload the video of the performance and metadata
that is necessary for the analysis. (c) Performance
feedback. The user shall be presented with feedback
about their performance.
3.2 Technical Design of the Application
The application is deployed inside of Microsoft’s
Azure cloud environment (Microsoft, 2023a) and uses
Azure resources for storing data (Azure blob storage)
and orchestrating the analysis (Azure container in-
stances).
The two backend services that process the input
video and calculate the analysis results are written
in Python. This was chosen, because the initial
processing of the input video is done using the
Python version of the Openpose system (Cao et al.,
2017).
Deployment: The chosen modular design of the ar-
chitecture allowed for custom deployment strategies
fitting the needs of the individual parts of the system
(Salah et al., 2016).
The python services were deployed inside short-
lived docker containers. This was done, because of
the high hardware requirements for the Openpose sys-
tem (Openpose, 2023), and this way the usage of the
hardware is limited to the time the service is running.
3.3 User Instructions
Depicted in Figure 1 is the examplary analysis that
is presented to the user when the application is first
opened. This provides the user with an insight how
the performance should be recorded and how the
technique should be executed. The performer in
the video is obfuscated and the analysis results are
blurred due to data privacy and ethical reasons.
2Trax3: Raising Accessibility and Everyday Use of Automatic Motion Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation
Algorithms
131
Figure 1: Analysis example presented to the user.
3.4 User Input
Figure 2 illustrates the interface where users can up-
load videos of performances. These videos capture
a person executing the double side kick technique,
recorded from a side view. Within the system, users
have the capability to upload both the video itself
and accompanying metadata regarding the performer.
This metadata is crucial for analysis purposes and in-
cludes details such as the leg used by the performer
for the kick and the performer’s height in centimeters,
as depicted in the video.
Figure 2: Upload of video including user metadata.
3.4.1 Analysis Algorithm
The analysis algorithm is a five-step process: (1)
coordinates for keypoints of the performer’s body are
detected and extracted from every frame of the video,
(2) the time series created by the set of coordinates
are cleaned by running them through an anomaly
detection, (3) 3D angles are calculated from the
extracted keypoints, (4) the calculated angles are
processed through an anomaly detection, (5) the
cleaned coordinates and angles are used to generate
the phases and nodes, as well as to extract and
calculate the performance indicators.
Image Processing: The frames of the video are pro-
cessed using the Openpose system (Cao et al., 2017).
Openpose takes a single frame as input, detects the
performer in the video and outputs coordinates for a
set of keypoints. The pose detection model is trained
on the COCO data set (Lin et al., 2014), which con-
tains 250.000 images of people each labeled with 18
keypoints of the human body. The coordinates of the
left and right shoulder, the left and right hip, the left
and right knee, the left and right ankle and the nose
of the performer are used in the following procedures.
Anomaly Detection: The anomaly detection com-
prises a set of predefined rules that marks output of
the previous step as anomalies that are then not con-
sidered for the calculation. It uses a time series as
input, created by either the x-coordinates (horizontal)
or y-coordinates (vertical) of the kicking leg’s ankle,
knee or hip. In this context the coordinates refer to the
position in the 2D image currently being processed.
Two different type of errors occur in the output: (i)
The body part cannot be detected in the frame and the
coordinates are either missing or zero, (ii) the body
part position is falsely detected and the coordinates
are present but incorrect.
The first type of error is handled by using the av-
erage between the first preceding non-zero coordinate
and the next non-zero coordinate to interpolate the
missing value. The second type of error is handled
by the rule set. A rule takes a consecutive sequence
of coordinates as an input and returns a boolean
indicating that sequence represent an anomaly or
not. The rules check for differences in value between
successive points in the sequence or check the
monotony of the sequence.
Calculation Characteristics: The determination of
knee angles involves an initial estimation of the miss-
ing depth coordinates, which are not available in the
2D analysis as compared to the 3D analysis. To ob-
tain these coordinates, the distances between the knee
and hip, as well as the knee and ankle of the kicking
leg, are computed. These distances are then sorted in
ascending order, and the upper value corresponding to
the 95
th
percentile of each set is selected as the "real"
length of the knee and hip or knee and ankle. This es-
timation of the depth coordinate is calculated for each
frame of the video analysis.
icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support
132
The obtained value is utilized to construct a right
triangle, considering the detected positions of the
knee and ankle, for instance. This enables the esti-
mation of the depth coordinate. By doing so, a 3D
vector is generated, representing the line connecting
the knee to the ankle, as well as the line connecting
the knee to the hip. Subsequently, the angles between
these 3D vectors are calculated. Furthermore, the cal-
culated angle values undergo anomaly detection, em-
ploying a specific set of rules designed for this pur-
pose. The identified coordinates and calculated an-
gles are employed to detect the nodes that signify
the commencement and conclusion of distinct phases
within the technique. To determine the INI node, the
y-coordinates of the ankle of the kicking leg are ex-
amined, specifically by identifying the first monoton-
ically increasing sequence of a predetermined length.
Regarding the MT1 node, the previously described
simple definition is utilized, involving the highest y-
coordinate of the knee prior to the angle surpassing
110 degrees. For the KE1 node, the first monotoni-
cally decreasing sequence of angles subsequent to the
MT1 node is considered. The same respective rules
are applied for the detection of the MT2 and KE2
nodes.
To compute the velocity and distance characteris-
tics, the distance between the nose and ankle is uti-
lized. Similarly, the upper value of the 95
th
percentile
is employed to determine the true height of the per-
former, as provided by the user, and map it to the cor-
responding coordinates in the video. This allows for
accurate estimation and calculation of the performer’s
velocity and distance metrics.
3.5 Statistical Evaluation
Only 26 of the full 44 videos from the preliminary
study, were suitable (due to quality issues) for analy-
sis.
The differences between 3D and 2D measure-
ments were all prechecked for normal distribution by
Shapiro-Wilk test using a significance level of α =
0.05.
The Bland-Altman plots estimated the average
mean difference agreement between the 3D and 2D
analysis for the phase duration in CH1 (-0.04 s; LOA
-0.15 to 0.06), KI1 (0.04 s; LOA -0.05 to 0.12), CH2
(-0.04 s; LOA -0.16 to 0.08), and KI2 (-0.01 s; LOA
-0.13 to 0.12).
Agreement for the femur angles at the node MT1
(-0.59
; LOA -20.20 to 19.03), KE1 (4.92
; LOA -
23.47 to 33.31), MT2 (4.13
; LOA -13.64 to 21.89),
KE2 (-11.59
; LOA -43.94 to 20.77).
Agreement for the leg spreading angles at the node
MT1 (4.25
; LOA -14.47 to 22.97), KE1 (6.84
; LOA
-17.63 to 31.31), MT2 (-0.36
; LOA -16.04 to 15.33),
KE2 (-7.08
; LOA -40.09 to 25.92).
Agreement for the relative knee height at KHK1 (-
5.78; LOA -29.76 to 18.20), KI1 (-3.30; LOA -23.95
to 17.36), KHK2 (-2.04; LOA -32.06 to 27.99).
Agreement for the relative shoulder knee distance
DSK 1(0.48; LOA -11.75 to 12.71), DSK 2 (1.41;
LOA -21.46 to 24.27).
A positive value of the average mean difference
indicates that the 2D analysis measured larger values
for this characteristic, and a negative value indicates
the 3D analysis has measured larger values compared
to the 2D analysis. The limits of agreement repre-
sent the range within which approximately 95 % of
the differences between the 2D and 3D measurements
will fall.
4 DISCUSSION
Generally it can be stated, that the main functional-
ity of the application is suitable for the purpose and
includes all defined functional requirements. Fur-
thermore, all non-functional requirements were met.
However it must be noted, that not all videos could be
analyzed due to quality issues of the recording, imply-
ing that the uploads must have a certain standard for
further processing. Despite, these quality issues of the
source material, which led to exclusion of 18 videos,
the accuracy of the extracted data is rather high.
Furthermore based on the plots, the 2D measure-
ments do not have any significant systematic bias for
any of the characteristics compared to the 3D mea-
surements.
In particular, based on the results of the Bland-
Altman plots, the hypothesis can be accepted for per-
formance indicators (i) Leg spreading vector at node
MT2, (ii) phase duration of CH1 and CH2, (iii) rel-
ative should distance DSK1 and relative knee height
KI1 but must be declined for the remaining distance,
angle, duration performance indicators.
There are studies that report a higher accuracy
in the kinematic analysis of 2D video (Schurr et al.,
2017; Peebles et al., 2021). In those two studies how-
ever the performer in the video remains in the same
position of the screen throughout the analysis. They
are either performing a single-leg squat or running on
a treadmill. The main challenge of the analysis in this
study was the movement of the performer, particularly
because of the missing depth parameters. There has
been a study using the Kinect
(Microsoft Corpora-
tion, Redmond, Washington, United States of Amer-
ica) 2D sensor for kinematic analysis (Pfister et al.,
2Trax3: Raising Accessibility and Everyday Use of Automatic Motion Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation
Algorithms
133
2014), but the results indicate that the measurements
cannot reach the accuracy of a 3D motion capture sys-
tem.
This study used only one type of camera and one
viewing angle for the recordings, which is common in
comparable 2D and 3D video investigations (Schurr
et al., 2017; Peebles et al., 2021; Remedios and Fis-
cher, 2021), although it would be interesting to exam-
ine the effect of different recording setups and dif-
ferent recording equipment on the analysis quality.
Particularly because it would better resemble the real
world environment, as athletes using the automated
analysis in their day to day training will not always
have the same recording quality and setup.
Besides effort and operation simplity, the cost fac-
tor is essential for integration of video analysis in reg-
ular training. The analysis in this study relies on hard-
ware that has a graphics processing unit. This type
of hardware is generally more expensive(Microsoft,
2023b). Similar analysis can be performed on cheaper
hardware that only contains a central processing unit,
however this leads to a much longer analysis duration
(Liang et al., 2019), which would prevent providing
instant feedback to the athletes.
Largely the challenges faced in this study are
similar to other studies with same pose estimation
software setup (Remedios and Fischer, 2021), which
means 2D video pose estimation is not yet ready to
replace more complex and expensive capturing sys-
tems, but serves well for performance estimations in
everyday training.
4.1 Limitations
The videos used for validating the application, were
recorded in 24 frames per second. A higher frame
rate would allow for a higher accuracy in measuring
the characteristics. In addition, the recording setup
for all of the videos was similar. Only a subset of the
test data was able to be used for validation, due to
poor video quality of parts of the test data (recording
equipment from approx. 2005). The use of differ-
ent camera types, video quality and recording setups
should be explored in a subsequent study.
5 CONCLUSIONS
Based on the current findings, it can be inferred that
the system, although in a prototypic stage, demon-
strates its suitability for automated analysis of 2D
double side kick videos. Moreover, it offers a conve-
nient and cost-effective means of technique analysis,
catering to a wide range of users, helping to identify
weaknesses for targeted innovative training (Hoel-
bling et al., 2021a) (Hoelbling et al., 2020). How-
ever, it should be noted that the system’s effectiveness
and accuracy heavily rely on scientifically established
performance indicators, as well as the quality of the
source material. In light of these considerations, it
can be concluded that the system serves as a valuable
tool for various applications.
REFERENCES
Adams, W. C. (2015). Conducting semi-structured inter-
views. Handbook of practical program evaluation,
pages 492–505.
Ambro
˙
zy, T., Rydzik, Ł., edra, A., Ambro
˙
zy, D.,
Niewczas, M., Sobiło, E., and Czarny, W. (2020). The
effectiveness of kickboxing techniques and its relation
to fights won by knockout. Archives of Budo, 16.
Barris, S. and Button, C. (2008). A review of vision-based
motion analysis in sport. Sports Medicine, 38:1025–
1043.
Bland, J. M. and Altman, D. (1986). Statistical methods for
assessing agreement between two methods of clinical
measurement. The lancet, 327(8476):307–310.
Bland, J. M. and Altman, D. G. (1999). Measuring agree-
ment in method comparison studies. Statistical meth-
ods in medical research, 8(2):135–160.
Cao, Z., Simon, T., Wei, S.-E., and Sheikh, Y. (2017). Real-
time multi-person 2d pose estimation using part affin-
ity fields. In CVPR.
DeFroda, S. F., Sugimoto, D., Staffa, S. J., Bae, D. S., Shan-
ley, E., Thigpen, C. A., and Kriz, P. K. (2021). Relia-
bility of an observational biomechanical analysis tool
in adolescent baseball pitchers. International Journal
of Sports Physical Therapy, 16(6):1523.
Elliott, B. (1999). Biomechanics: an integral part of sport
science and sport medicine. Journal of Science and
Medicine in Sport, 2(4):299–310.
Floyd, C. (1984). A systematic look at prototyping. In
Approaches to prototyping, pages 1–18. Springer.
Göhner, U. (1992). Movement theory: Fundamental as-
pects. Sport Science in Germany: An Interdisciplinary
Anthology, pages 191–200.
Hoelbling, D., Grafinger, M., Baca, A., and Dabnichki, P.
(2020). The flexibility trainer: Feasibility analysis,
prototype-and test station development for a sports de-
vice for hip-joint flexibility and strength enhancement.
In 8th International Conference on Sport Sciences Re-
search and Technology Support (ICSPORTS), pages
22–29.
Hoelbling, D., Grafinger, M., Smiech, M. M., Cizmic, D.,
Dabnichki, P., and Baca, A. (2021a). Acute response
on general and sport specific hip joint flexibility to
training with novel sport device. Sports Biomechan-
ics, pages 1–16.
Hoelbling, D., Smiech, M. M., Cizmic, D., Baca, A., and
Dabnichki, P. (2021b). Exploration of martial arts kick
icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support
134
initiation actions and telegraphs. International Jour-
nal of Performance Analysis in Sport, 21(4):507–518.
Hölbling, D., Baca, A., and Dabnichki, P. (2020a). A
kinematic model for assessment of hip joint range-of-
motion in fast sport movements using spreading an-
gles. Sports Biomechanics, pages 1–13.
Hölbling, D., Baca, A., and Dabnichki, P. (2020b). Sequen-
tial action, power generation and balance characteris-
tics of a martial arts kick combination. International
Journal of Performance Analysis in Sport, 20(5):766–
781.
Hölbling, D., Preuschl, E., Hassmann, M., and Baca, A.
(2017). Kinematic analysis of the double side kick in
pointfighting, kickboxing. Journal of sports sciences,
35(4):317–324.
Lapham, A. and Bartlett, R. (1995). The use of artificial in-
telligence in the analysis of sports performance: A re-
view of applications in human gait analysis and future
directions for sports biomechanics. Journal of Sports
Sciences, 13(3):229–237.
Lees, A. (2002). Technique analysis in sports: a critical
review. Journal of sports sciences, 20(10):813–828.
Liang, X., Kapetanios, E., Woll, B., and Angelopoulou, A.
(2019). Real time hand movement trajectory track-
ing for enhancing dementia screening in ageing deaf
signers of british sign language. In Machine Learn-
ing and Knowledge Extraction: Third IFIP TC 5, TC
12, WG 8.4, WG 8.9, WG 12.9 International Cross-
Domain Conference, CD-MAKE 2019, Canterbury,
UK, August 26–29, 2019, Proceedings 3, pages 377–
394. Springer.
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,
Ramanan, D., Dollár, P., and Zitnick, C. L. (2014).
Microsoft coco: Common objects in context. In Com-
puter Vision–ECCV 2014: 13th European Confer-
ence, Zurich, Switzerland, September 6-12, 2014, Pro-
ceedings, Part V 13, pages 740–755. Springer.
Maykut, J. N., Taylor-Haas, J. A., Paterno, M. V., DiCe-
sare, C. A., and Ford, K. R. (2015). Concurrent valid-
ity and reliability of 2d kinematic analysis of frontal
plane motion during running. International journal of
sports physical therapy, 10(2):136.
Mayring, P. et al. (2004). Qualitative content analysis. A
companion to qualitative research, 1(2):159–176.
Meinel, K. and Schnabel, G. (2007). Bewegungslehre-
Sportmotorik: Abriss einer Theorie der sportlichen
Motorik unter pädagogischem Aspekt. Meyer &
Meyer Verlag.
Microsoft (2023a). Azure cloud computing services.
https://azure.microsoft.com/en-gb/. Accessed: 2023-
05-31.
Microsoft (2023b). Azure machine pricing. https://
azure.microsoft.com/en-us/pricing/details/virtual-
machines/windows/. Accessed: 2023-06-01.
Openpose (2023). Openpose hardware requirements.
https://github.com/tramper2/openpose/blob/master/d
oc/installation.md. Accessed: 2023-06-01.
Ouergui, I., Delleli, S., Bouassida, A., Bouhlel, E.,
Chaabene, H., Ardigò, L. P., and Franchini, E. (2021).
Technical-tactical analysis of small combat games in
male kickboxers.
Peebles, A. T., Carroll, M. M., Socha, J. J., Schmitt, D.,
and Queen, R. M. (2021). Validity of using automated
two-dimensional video analysis to measure contin-
uous sagittal plane running kinematics. Annals of
biomedical engineering, 49:455–468.
Pfister, A., West, A. M., Bronner, S., and Noah, J. A. (2014).
Comparative abilities of microsoft kinect and vicon 3d
motion capture for gait analysis. Journal of medical
engineering & technology, 38(5):274–280.
Remedios, S. M. and Fischer, S. L. (2021). Towards the use
of 2d video-based markerless motion capture to mea-
sure and parameterize movement during functional
capacity evaluation. Journal of Occupational Reha-
bilitation, 31:754–767.
Salah, T., Zemerly, M. J., Yeun, C. Y., Al-Qutayri, M., and
Al-Hammadi, Y. (2016). The evolution of distributed
systems towards microservices architecture. In 2016
11th International Conference for Internet Technology
and Secured Transactions (ICITST), pages 318–325.
IEEE.
Schurr, S. A., Marshall, A. N., Resch, J. E., and Saliba, S. A.
(2017). Two-dimensional video analysis is compara-
ble to 3d motion capture in lower extremity movement
assessment. International journal of sports physical
therapy, 12(2):163.
Yuncza, E. (2011). Sport karate point sparring: An essen-
tial guide to the point fighting method. Indomitable
Pub.
Zhang, C. and Lu, Y. (2021). Study on artificial intelligence:
The state of the art and future prospects. Journal of
Industrial Information Integration, 23:100224.
Zorzetti, M., Signoretti, I., Salerno, L., Marczak, S., and
Bastos, R. (2022). Improving agile software devel-
opment using user-centered design and lean startup.
Information and Software Technology, 141:106718.
2Trax3: Raising Accessibility and Everyday Use of Automatic Motion Analysis in (Combat) Sports via ML Enhanced 2D to 3D Estimation
Algorithms
135