Deep Learning-Based Autoencoder for Objective Assessment of

Taekwondo Poomsae Movements

Mohamed Amine Chaâbane

, Imen Ben Said

1,2 b

and Adel Chaari

MIRACL Laboratory, University of Sfax, Tunisia

Digital Research Center of Sfax, Technopark of Sfax, PO Box 275, 3021 Sfax, Tunisia

hejAL! New Tech Solutions GmbH, Bad Dürkheim, Germany

Keywords: Poomsae Movements, Assessment, Deep Learning, Autoencoder, Skeleton Body Points, Computer Vision.

Abstract: Artificial Intelligence (AI) is revolutionizing sports by enhancing performance, improving safety, and creating

richer fan experiences. This paper focuses on leveraging AI in Taekwondo, specifically in assessing athlete

performance during Poomsae movements, which are foundational to the sport and crucial for success in

competitions. Traditionally, the evaluation of Poomsae has been subjective and heavily reliant on human

judgment. This study addresses this issue by automating the assessment process. We propose a deep learning

approach that utilizes computer vision to analyse athletes' movements captured in video clips of Poomsae.

The proposed approach is based on a model that emphasizes the use of autoencoders for training data

representing skeleton body points of correct movements. This model can effectively identify anomalies, i.e.,

incorrect movements by athletes. The SportLand platform implements the proposed approach, providing

coaches and athletes with precise and actionable insights into their performance. This platform can serve as

an assistant for self-evaluation, allowing Taekwondo athletes to enhance their skills at their own pace.

1 INTRODUCTION

Nowadays, the use of AI-driven systems in sports,

including machine learning (ML), computer vision,

and data analytics, has led to a significant paradigm

shift in athlete management. These systems transform

the landscape of sports science, significantly

supporting real-time decision-making in coaching

(Pisaniello, 2024). Recently, integrating AI

technologies to enhance athlete performance has

become a key focus of research (Hong et al., 2021;

Michalski et al., 2022).

Taekwondo is a widely practiced martial art, with

Poomsae as a fundamental component that consists of

choreographed patterns embodying the core

techniques and principles of the art. Traditionally,

Poomsea movement evaluation relies mainly on

subjective visual observation, which can lead to

errors and bias in judgment. As a result, there is a

growing demand for automated solutions that offer

effective and objective methods to assess the quality

and accuracy of athletes' Poomsae movements.

https://orcid.org/0000-0002-6831-5530

https://orcid.org/0000-0001-5025-6715

Computer vision has emerged as a promising

solution to address the subjectivity often associated

with manual evaluation methods. Specifically, human

action/activity recognition (HAR) plays a crucial role

in identifying and labelling activities from videos that

capture complete actions (Kong & Fu, 2022) by

utilizing AI and various hardware devices, such as

cameras. Due to its cost-effectiveness, Sun et al.

(2022) argue that skeleton point detection is a more

effective data modality for HAR than other

modalities, such as red-green-blue (RGB) depth

images. The authors also assert that deep learning DL

methods, such as convolutional neural networks

(CNNs), autoencoders, and long short-term memory

(LSTM) networks, have demonstrated significant

potential in HAR tasks.

HAR has been widely researched in sports, with its

use to analyze and evaluate various athletic performan-

ces (Host & Ivašić-Kos, 2022). This paper proposes an

AI-based approach to assessing Taekwondo

movements, thereby avoiding the subjectivity inherent

in traditional evaluation methods. This approach can

170

Chaâbane, M. A., Ben Said, I. and Chaari, A.

Deep Learning-Based Autoencoder for Objective Assessment of Taekwondo Poomsae Movements.

DOI: 10.5220/0013706100003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 170-177

ISBN: 978-989-758-771-9; ISSN: 2184-3201

significantly enhance athlete performance by

providing real-time insights into movement accuracy

and kinematic features, such as force and speed.

More precisely, the proposed approach leverages

the power of deep learning autoencoder models and

utilizes variations in skeleton point data to evaluate

Poomsae movements, effectively distinguishing

between correct and incorrect movements. To achieve

this, we created a dataset of videos showcasing the

correct execution of Poomsae by skilled athletes. This

dataset was then used to train and evaluate the

autoencoder model, which achieved an impressive

average accuracy of 99%.

The SportLand platform is a specific tool that

implements our DL approach. It assists athletes in

enhancing their performance by offering a self-

training service designed to sharpen their skills.

The remainder of this paper is organized as

follows: Section 2 provides a comprehensive

overview of the general context and prior research in

the field, delving into the background and existing

literature. Following this, Section 3 outlines our deep

learning approach for evaluating Poomsae. Section 4

introduces the SportLand platform, designed for

delay training to enhance athletes' performance.

Finally, Section 5 summarizes the study's key

findings and offers directions for future work.

2 BACKGROUND AND RELATED

WORKS

2.1 Poomsae in Taekwondo Sports

In Taekwondo, a poomsae is a sequence of basic

movements that encompasses both offensive and

defensive techniques suitable for competition. Figure

1 depicts the movements of Taegeuk I Jang, one of

the foundational poomsae in Taekwondo. This form

includes essential actions such as walking and basic

techniques like Makki (block) and Chagi (kick).

Taegeuk I Jang consists of 18 movements, numbered

from 1 to 18, as shown in Figure 1.

Furthermore, to characterize a Poomsae movement

as correct, the World Taekwondo Federation (WTF,

2014) provides a set of guidelines, including:

• Pause Between Movements: Athletes should

incorporate a brief pause between movements to

emphasize control and precision, allowing for a

clear distinction between each movement.

• Symmetrical Pattern of Poomsae Line: The

Poomsae should follow a symmetrical pattern,

ensuring that movements are executed evenly on

both sides, reflecting the balance and harmony

inherent in Taekwondo.

• Balance of Each Movement: Maintaining balance

throughout each movement is crucial, as it ensures

stability and effectiveness in techniques, allowing

for powerful and controlled execution.

Figure 1: Taegeuk I Jang Poomsae movements.

2.2 Athlete Performance Assessment in

Literature

According to the literature review based on HAR for

martial arts performance evaluation, several existing

approaches utilize video analysis to assess athletes'

movements based on skeleton points data.

A paper by Lee and Jung (2020) introduces a

reliable Taekwondo Poomsae movement dataset

called TUHAD and proposes a key-frame-based

CNN architecture for recognizing Taekwondo actions

using this dataset.

Barbosa et al. (2021) compare four different deep

learning models to classify Taekwondo movements,

aiming to identify which model yields the best results.

The study found that convolutional layer models,

including CNN combined with LSTM and

Convolutional Long Short-Term Memory

(ConvLSTM) models, achieved over 90% accuracy in

classification.

Emad et al. (2020) propose a smart coaching

system called iKarate for Karate training, which tracks

players' movements using an infrared camera sensor.

After a preprocessing phase, the system classifies the

data using the fast dynamic time warping algorithm. As

a result, the proposed system generates a detailed

report outlining each action performed by the player,

identifying mistakes in every movement, and

providing suggestions for improvement.

Deep Learning-Based Autoencoder for Objective Assessment of Taekwondo Poomsae Movements

171

Fernando et al. (2023) evaluate Taekwondo

movements using an LSTM-based machine learning

model, with a focus on classification performance. In

their evaluation, the LSTM model achieved an

accuracy of 96% on the test dataset and 61% on the

validation dataset.

After a thorough analysis of the literature review,

several limitations within the current systems have

been identified. On one hand, the proposed

approaches often involve costly logistical

preparations, such as specialized hardware (sensors,

infrared trackers, Kinect cameras) or dedicated

laboratory environments. On the other hand, these

contributions primarily focus on classifying

sequences of actions using classification models to

evaluate athletes' movements. While these models are

well-suited for sequence-based tasks and effectively

leverage their memory capabilities to learn from time-

dependent data, they are less effective at learning

efficient data representations. This is crucial for

accurately evaluating movements and distinguishing

between correct and incorrect actions.

3 AUTOENCODER FOR

POOMSAE MOVEMENTS

ASSESSMENT

This section outlines the deep learning approach we

propose for assessing athlete performance. Figure 2

illustrates the general process for implementing our

deep learning approach to evaluate Poomsae sequences.

In the following subsections, we will detail each

step of this process.

3.1 Data Collection

The Data Collection step aims to create the dataset that

the proposed model uses. To do so, we collect Poomsae

sequences from diverse online sources. This includes

videos available on platforms such as YouTube and

martial arts websites that are carefully chosen to cover

a comprehensive range of movements necessary for

Poomsae sequences. In addition, we record video

sequences during training sessions supervised by

renowned professional trainers. These videos provide

examples of movements performed by skilled athletes

and offer high-quality data. Specific conditions, such

as lighting and camera angles, were adhered to while

MediaPipe is a robust and highly accurate open-source

framework for real-time pose detection, which allows for

precise identification of the human body's critical

anatomical landmarks.

recording the videos, ensuring the consistent and

superior capture of the athletes' movements.

Figure 2: Overview of the Steps Involved in Implementing

the Autoencoder approach for Assessing Athlete

movement.

3.2 Data Pre-Processing

Data Preprocessing is crucial in transforming and

preparing the data collected for effective utilization

by the autoencoder model. It involves the following

steps, which are detailed in the following sub-

sections: Frame Extraction, Body Point Detection,

Angle Calculation, Data Storage and Augmentation,

and Data Scaling.

3.2.1 Frame Extraction

Each collected video was manually segmented into

short video clips containing only one movement.

Then, key frames were extracted from the segmented

videos to define representative images of each

movement in the Poomsae sequences.

It is mandatory to bring all the video clips into the

same frame rate during extraction. More precisely,

the approach involved using a video processing

library to standardize the frame rate and duration of

the video clip. Specifically, the frame rate was

adjusted to 30 frames per second (fps) with a

consistent one-second duration per video clip. This

process ensured temporal uniformity, facilitating

precise and systematic analysis of the video content,

which allows for the precise identification and

analysis of specific movements.

3.2.2 Body Point Detection

Once the video frames are extracted, we leverage the

MediaPipe

library to detect and map each frame's

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

172

key body joints and limb positions. Figure 3 shows an

example of a body landmark driven by the MediaPipe

library. From this figure, we highlight that MediaPipe

provides a total of 33 landmarks for the human body,

including major joints and body parts, such as

shoulders, elbows, wrists, hips, knees, and ankles.

Figure 3: Example of Body Landmark Visualization driven

by MediaPipe library.

3.2.3 Angle Calculation

We have developed an algorithm to calculate precise

angles between various body landmarks for each

extracted frame. For example, in Figure 4, parts (a)

and (b) illustrate the angle BÂC, representing the

angle between the right arm and forearm. Here, points

A, B, and C correspond to landmarks 13, 11, and 15,

respectively (Cf. Figure 3). Part (c) of the figure

depicts the angle FÊG, which represents the angle

between the right upper leg and lower leg, with points

E, F, and G corresponding to landmarks 25, 23, and

27, respectively.

3.2.4 Data Storage and Augmentation

The calculated angles serve as features that describe

the input variables from which our autoencoder

model learns. These angles were stored in a CSV file,

ensuring easy access and facilitating in-depth analysis

of the athlete's kinematic information. Additionally,

this dataset can be utilized to train and evaluate deep

learning models for automated movement analysis

and assessment.

Once the initial CSV dataset is created, it is

important to enhance the stored kinematic data using

data augmentation techniques. Specifically, we

Figure 4: Example of Calculated Angles.

employ linear interpolation to generate additional

data points between each pair of adjacent rows in the

dataset. More precisely, we propose to add Gaussian

noise to support small random variations to our

numerical data. This is accomplished by adding

values drawn from a normal distribution (Gaussian

distribution) centred around zero to each data point.

This augmentation process increases the overall

size and diversity of the dataset, which can improve

the performance of any DL models trained on this

data.

3.2.5 Data Scaling

To ensure consistency across the dataset and

improve the performance of any DL model trained

on this data, we scaled the athlete kinematic features

using the MinMaxScaler module from the scikit-

learn preprocessing library. This scaling technique

normalizes the data to a common numerical range,

typically between 0 and 1. This data normalization

process helps to achieve better convergence and

stability during the training of deep learning models,

as it ensures that all features are on a similar scale

and have a comparable influence on the learning

process.

3.3 Data Splitting

The dataset was divided into training and testing sets.

We typically use 70% of the data for training and 30%

for testing. This division enables the model to acquire

knowledge from most of the data while being

evaluated on a subset.

3.4 Model Training

3.4.1 Model Building

The deep learning approach we propose leverages

autoencoders to detect incorrect movements of

athletes. Specifically, we define and train an

autoencoder model on data representing correct

movements, enabling it to reconstruct athlete

movements accurately.

Deep Learning-Based Autoencoder for Objective Assessment of Taekwondo Poomsae Movements

173

The architecture of the proposed autoencoder

model is given in Figure 5. The autoencoder contains

three main components: the Encoder, the latent space,

and the Decoder. The architecture diagram illustrates

the flow from input to output through these

components. More precisely, the Encoder component

encompasses the following layers:

 Input Layer: The model begins with an input

layer that accepts data with features

corresponding to calculated angles.

 First Dense Layer (256): The input data is

processed through a dense layer of 256

neurons. This layer applies L2 regularization

with a coefficient of 0.01, followed by batch

normalization to stabilize and accelerate the

training process. LeakyReLU's activation

function introduces non-linearity while

allowing a small gradient when the unit is not

active.

 First Dropout Layer (0.3): A dropout layer

with a rate of 0.3 is applied to randomly set

30% of the input units to zero during training.

This will further help to prevent overfitting.

 Second Dense Layer (128): This is followed

by a dense layer with 128 neurons, which

again applies L2 regularization, batch

normalization, and LeakyReLU activation.

 Second Dropout Layer (0.3): A second

dropout layer with a rate of 0.3 is included.

 Third Dense Layer (64): The final layer of the

encoder has 64 neurons, with the same

regularization and activation techniques as the

previous layers. This reduces the input to a

smaller dimensional space.

Latent Space (code representation): The latent

space representation captures the essential features of

the input data in a lower-dimensional form.

The decoder is composed of the following layers:

 First Dense Layer (128): The latent space is

redirected through a dense layer comprising

128 neurons, utilizing the identical L2

regularization, batch normalization, and

LeakyReLU activation.

 Second Dense Layer (64): The next layer

further decreases the dimensionality with 64

neurons, continuing with regularization and

activation techniques.

 Final Dense Layer (4): The final dense layer

contains 4 neurons, matching the input

dimension. It uses the sigmoid activation

function to ensure that the output values are

normalized between 0 and 1.

Figure 5: Architecture of the Autoencoder model.

3.4.2 Reconstruction Algorithm

The reconstruction algorithm's first step is to

decompose the Poomsae video into individual

movements for evaluation. The algorithm receives an

initial set of 30 frames representing one movement,

extracts key points for each frame, and calculates the

corresponding angles. All calculated angles, along

with their respective frame numbers, are then input

into the autoencoder model. The model subsequently

performs the reconstruction task. This process is

repeated for the entire video.

3.4.3 Model Compilation

The autoencoder model is compiled using the Adam

optimizer along with a Mean Squared Error (MSE)

loss function. The Adam optimizer was chosen for its

efficient performance with large datasets and its

adaptive learning rate mechanism, which adjusts the

learning rate for each parameter. The MSE loss

function is particularly well-suited for regression

tasks, as it minimizes the reconstruction error. This

minimization is crucial for effective anomaly

detection, allowing the model to accurately

reconstruct normal data patterns and identify

deviations representing incorrect movements.

Furthermore, to prevent overfitting and optimize

training, we implemented the following callbacks:

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

174

 EarlyStopping: This callback monitors the

validation loss during training and stops the

training process if the validation loss does not

improve for a specified number of epochs. This

helps to avoid overfitting and ensures that the

model maintains its best performance.

 ModelCheckpoint: The model with the lowest

validation loss is saved by this callback. The

saved model can then be used for further

evaluation and testing.

3.4.4 Training Results

The autoencoder models were trained using the

training dataset with the following parameters: 100

epochs, a batch size of 32, and a learning rate of

0.002. The model was trained with the objective of

minimizing the reconstruction error. Furthermore, to

ensure the effectiveness of the training process, the

loss and validation loss performance measures were

monitored and plotted over the epochs. This revealed

how well the model was learning and whether it was

over-fitting or under-fitting.

Figure 6 illustrates the training results of the first

four movements of Taegeuk I Jang Poomsae. The

interpretation of the results is as follows: Initially,

both training and validation losses went down quickly

in the beginning, which showed that the model was

learning well from the input data. The loss error

stabilized and reached near-zero values, showing that

the models could reconstruct the input data with

minimal error. The best epoch, marked by the lowest

validation loss, was identified for each movement,

making sure that optimal weights were saved for later

evaluation.

Figure 6: Training and Validation Loss Over Epochs for

three different movements of Taegeuk I Jang Poomsae.

https://www.sportland.ai

The overall training results were highly

satisfactory, with each movement achieving low

reconstruction errors and demonstrating robust

generalization capabilities. The model's ability to

accurately reconstruct the input data while

maintaining low validation loss highlights its

effectiveness in capturing the correct patterns for each

movement in Taegeuk I Jang Poomsae.

3.5 Model Evaluation

The model's effectiveness is assessed using the Mean

Squared Error (MSE) and the coefficient of

determination (R2) for both the training and testing

datasets. Results for the first three movements related

to Taegeuk I Jang Poomsae are summarized in Table

1, presented below.

According to the presented results, the train and

test MSE values are consistently low, indicating that

the model performs well on both training and test

datasets. The Train and Test R² values are close to 1,

indicating that the model explains a large proportion

of the variance in the data, both for training and

testing. The performance metrics across different

movements are consistent, indicating that the model

can adapt to various types of movements.

Table 1: Results of model evaluation.

Taegeuk I

Jang

movements

Train

MSE

Test

MSE

Train R² Test R²

Movement1 0.00025 0.00025 0.99451 0.99442

Movement2 0.00026 0.00026 0.99225 0.99216

Movement3 0.00022 0.00023 0.99464 0.99445

4 SPORTLAND PLATFORM

SportLand

is a sports tech platform that connects

athletes across various disciplines, particularly in

martial arts such as Taekwondo, karate, kung fu, and

kickboxing. The platform features SportLand

AICoach

module, which is an AI-powered tool designed to

evaluate athlete performance. It analyses an athlete's

movements during training, providing actionable

insights to optimize technique and maximize

performance gains. This tool serves as a self-

evaluation assistant, enabling Taekwondo athletes to

enhance their skills at their own pace.

Deep Learning-Based Autoencoder for Objective Assessment of Taekwondo Poomsae Movements

175

More precisely, the SportLand

AICoach

module

offers two distinct methods for evaluating athlete

performance: Single-Movement Evaluation and Full-

Sequence Evaluation

The next subsections provide detailed

information on how these evaluations are conducted

within the SportLand platform. We also provide an

online demonstration video

illustrating these

evaluations.

4.1 Single-Movement Evaluation

The SportLand

AICoach

platform provides a

comprehensive interface for uploading, displaying,

and analyzing videos, offering real-time feedback on

movement correctness and instantaneous speed and

force metrics. Athletes can upload pre-recorded

videos or use their webcam to capture live footage.

Once a video is uploaded or recorded, it is

decomposed into 30 frames representing the

movement to be evaluated. Next, body landmarks are

extracted, and angles are calculated from each frame.

All calculated angles and their respective frame

numbers are input into the trained model to compute

the reconstruction error. If this error exceeds a

predetermined threshold, the movement is flagged as

incorrect. Conversely, if the error is within acceptable

limits, the movement is deemed correct, and the

algorithm subsequently calculates the athlete's speed

and force. Figure 7 illustrates an example of a

correctly evaluated movement.

4.2 Full-Sequence Evaluation

SportLand

AICoach

offers the ability to analyze the

complete sequence of movements in a Poomsae. To

achieve this, the athlete must upload a video, like the

method used for evaluating individual movements.

The uploaded video is then broken down into

individual movements for assessment, and the

previously described steps for assessing a single

movement are applied to all segments.

The athlete’s score increases with each correctly

identified movement, while incorrectly identified

movements receive 0 points. Additionally, performing

the movements in the correct order is essential, as each

movement must accurately follow its predecessor. A

movement is marked as incorrect if it is not detected

within 60 frames, ensuring timely execution and

thorough evaluation. Each correct movement adds 2

points to the athlete's score, while incorrect movements

or those performed out of sequence score 0 points. The

https://www.youtube.com/watch?v=cOXxqcG1v8A

Figure 7: Example of evaluation result.

total score, which reflects the athlete's accuracy and

performance, is calculated by summing the individual

scores for each movement.

Figure 8 below illustrates the interface displaying

a Poomsae score. The interface provides athletes with

their total score and features a comprehensive table

detailing each movement performed. This table

displays the results for every individual movement,

indicating whether each movement was assessed as

correct or incorrect, along with the corresponding

score for each movement.

5 CONCLUSION

The primary research problem addressed in this study

was the subjectivity inherent in traditional methods of

evaluating Taekwondo Poomsae. This paper

proposed a deep learning approach based on an

autoencoder model, which effectively mitigates

human intervention and demonstrates the potential

for a systematic evaluation of Poomsae.

The proposed autoencoder model identified

anomalies in an athlete's movements by analyzing a

sequence of frames and considering movement

behavior through skeleton point data, rather than

focusing solely on the final pose. Additionally, the

model was tested on a diverse set of video data to

evaluate its generalizability to real-world scenarios.

In this assessment, the autoencoder exhibited

remarkable performance, achieving a coefficient of

determination (R²) close to 99% and a very low Mean

Squared Error (MSE) of approximately 0.0006.

Furthermore, the absence of self-evaluation

methods was identified as another challenge this

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

176

Figure 8: Interface displaying the total score and detailed

results for each Poomsae movement.

study aimed to address. To resolve this issue, we

successfully developed the SportLand

AICoach

, a

reliable and objective self-paced evaluation tool. This

tool leverages the autoencoder model, significantly

advancing the field of Poomsae evaluation.

While this study primarily focused on the

accuracy of Poomsae movements, future research

could extend to evaluating player movements during

Taekwondo combat. The goal of this evaluation

would be to assist judges in more accurately

computing athletes' scores during matches. By

leveraging skeleton point data, researchers could

analyze movement patterns in real-time to ensure that

scoring reflects the precision and effectiveness of

techniques used in combat. This approach could

enhance the fairness and accuracy of scoring and

contribute to a deeper understanding of performance

dynamics in competitive Taekwondo.

REFERENCES

Barbosa P., Cunha P., Carvalho V. & Soares F. (2021).

Classification of taekwondo techniques using deep

learning methods: First insights. Proceedings of the

BIODEVICES 2021 - 14th International Conference on

Biomedical Electronics and Devices; Part of the 14th

International Joint Conference on Biomedical

Engineering Systems and Technologies, BIOSTEC

2021, 11–13 January, Vienna, Austria, pp. 201–208.

Emad B., Atef O., Shams Y., El-Kerdany A., Shorim N.,

Nabil A. & Atia A. (2020). Ikarate: Karate Kata

guidance system. Procedia Computer Science

175(2019): 149–156. DOI: https://doi.org/10.1016/j.

procs.2020.07.024

Fernando, M., Sandaruwan, K. D., & Athapaththu, A. M.

K. B. (2024). Evaluation of Taekwondo Poomsae

movements using skeleton points.

Hong, S., Park, J., Lim, J. (2021). AI-powered wearables

for real-time performance feedback: A study in cycling.

Sports Technology Journal, 15(1), 45-57.

Host K. & Ivašić-Kos M. (2022). An overview of human

action recognition in sports based on computer vision.

Heliyon 8(6): e09633. DOI: https://doi.org/https://doi.

org/10.1016/j.heliyon.2022. e09633.

Kong Y. & Fu Y. (2022). Human action recognition and

prediction: a survey. International Journal of Computer

Vision 130(5): 1366–1401.

Michalski, R., Jones, T., Liu, C. (2022). Understanding AI

model transparency in sports analytics. Journal of

Applied Sports Science, 10(3), 95-108.

Pisaniello, A. (2024). The Game Changer: How Artificial

Intelligence is Transforming Sports Performance and

Strategy. Geopolitical, Social Security and Freedom

Journal, 7(1), 75-84.

Sun Z., Ke Q., Rahmani H., Bennamoun M., Wang G. &

Liu J. (2022). Human action recognition from various

data modalities: A Review. IEEE Transactions on

Pattern Analysis and Machine Intelligence 45(3):

3200–3225. DOI: https://doi.org/https://doi.org/10.

1109/tpami.2022.31 83112.

WTF (2014). Poomsae Scoring guidelines for International

Referees. World Taekwondo Federation. Available at

https://d17nlwiklbtu7t.cloudfront.net/983/document/

Poomsae_scoring_guidelines.pdf.

Deep Learning-Based Autoencoder for Objective Assessment of Taekwondo Poomsae Movements

177