Novasinergia 2026, 9(2), 148-173. https://doi.org/10.37135/ns.01.18.08 http://novasinergia.unach.edu.ec
Research article
Machine learning-based classification of push techniques in speed skating
Clasificación basada en aprendizaje automático de técnicas de empuje en patinaje de
velocidad
Ximena Albornoz-Tepan1, Sebastián Ulloa-Montaleza1, Jorge Barreto2,
Luis I. Minchala1, Fabian Astudillo-Salinas1
1Departamento de Ingeniería Eléctrica, Electrónica y de Telecomunicaciones, Universidad de Cuenca, Cuenca, Ecuador, 010207;
2Carrera de Pedagogía de la Actividad Física y el Deporte, Universidad de Cuenca, Cuenca, Ecuador, 010203;
ximena_alb@hotmail.com; bryan.ulloa.cuenca@gmail.com; jorge.barreto@ucuenca.edu.ec; isamel.minchala@ucuenca.edu.ec
*Correspondencia: fabian.astudillos@ucuenca.edu.ec
Citación: Albornoz-Tepan, X.;
Ulloa-Montaleza, S.; Barreto, J.;
Minchala, L. & Astudillo-Salinas,
F., (2026). Machine Learning-
Based Classification of Push
Techniques in Speed Skating.
Novasinergia. 9(2). 148-173.
https://doi.org/10.37135/ns.01.18.
08
Recibido: 04 marzo 2026
Aceptado: 28 mayo 2026
Publicado: 08 julio 2026
Novasinergia
ISSN: 2631-2654
Abstract: Speed skating is a prestigious sport that requires technical skills and optimal physical
condition. During training, computer vision and Machine Learning (ML) techniques can help
improve skating performance and biomechanical analysis. However, limited research has
addressed the automated classification of speed skating push techniques using pose estimation
methods. In this regard, the OpenPose model is used to extract data on the skater’s joints and key
points for movement analysis and push classification. The classification methodology implied
exploring two main approaches. The first approach uses image classification based on Skeleton Gait
Energy Images (SGEI) and a Convolutional Neural Network (CNN) with the VGG19 architecture
and transfer learning, achieving an accuracy of 90.72%. The second approach uses biomechanical
feature vectors through a Support Vector Machine (SVM) system and a Random Forest (RF)
algorithm, achieving accuracies of 94% and 92%, respectively. The classification task considered
three skating push techniques: classic push, double push, and pendulum push, which are
biomechanically relevant due to their influence on propulsion efficiency, balance, and skating
performance. Key findings indicate that feature-based models (SVM/RF) achieved higher precision
and faster execution, while the CNN approach provided greater flexibility through data
augmentation and automated parameter tuning. Furthermore, the “double push” technique was
the most accurately classified movement across all evaluated models. The proposed framework
contributes to the limited research on automated classification of skating push using pose
estimation and ML techniques for biomechanical and sports performance analysis.
Keywords: Biomechanics, Convolutional Neural Network, Machine Learning, OpenPose, Speed
skating.
Copyright: 2026 derechos otorgados
por los autores a Novasinergia.
Este es un artículo de acceso abierto
distribuido bajo los términos y
condiciones de una licencia de
Creative Commons Attribution (CC
BY NC).
(http://creativecommons.org/licenses/
by-nc/4.0/).
Resumen: El patinaje de velocidad es un deporte prestigioso que requiere habilidades técnicas y una condición
física óptima. Durante el entrenamiento, las técnicas de visión artificial y aprendizaje automático (ML)
pueden contribuir a mejorar el rendimiento en el patinaje y el análisis biomecánico. Sin embargo, la
investigación sobre la clasificación automatizada de las técnicas de empuje en el patinaje de velocidad mediante
métodos de estimación de la postura es limitada. En este sentido, se propone el modelo OpenPose para obtener
datos de las articulaciones y puntos clave del patinador para el análisis del movimiento y la clasificación del
empuje. La metodología de clasificación implicó explorar dos enfoques principales. El primer enfoque utiliza
la clasificación de imágenes mediante imágenes de energía de la marcha del esqueleto (SGEI) y una red
neuronal convolucional (CNN) basada en la arquitectura VGG19 con aprendizaje por transferencia, logrando
una precisión del 90,72 %. El segundo enfoque utiliza vectores de características biomecánicas mediante un
sistema de máquina de vectores de soporte (SVM) y un algoritmo de bosque aleatorio (RF), logrando
precisiones del 94 % y el 92 %, respectivamente. La tarea de clasificación consideró tres técnicas de impulso
en patinaje: impulso clásico, impulso doble e impulso pendular, las cuales son biomecánicamente relevantes
debido a su influencia en la eficiencia de la propulsión, el equilibrio y el rendimiento en patinaje. Los hallazgos
clave indican que los modelos basados en características (SVM/RF) lograron mayor precisión y una ejecución
más rápida, mientras que el enfoque CNN proporcionó mayor flexibilidad a través del aumento de datos y el
ajuste automático de parámetros. Además, la técnica de "impulso doble" fue el movimiento clasificado con
mayor precisión en todos los modelos evaluados. El marco propuesto contribuye a la escasa investigación sobre
la clasificación automatizada del impulso en patinaje mediante la estimación de la pose y técnicas de
aprendizaje automático para el análisis biomecánico y del rendimiento deportivo.
Palabras clave: Biomecánica, Red neuronal convolucional, Aprendizaje automático, OpenPose, Patinaje de
velocidad.
Novasinergia 2026, 9(2), 148-173 149
1. Introduction
Speed skating is a sport in which mastering a specific technique is crucial for success
[1]. Nowadays, multimedia technologies are employed to analyze images and study sports
videos to refine athletes' techniques [2]. For instance, the Kinovea software facilitates the
analysis of general movement biomechanics by examining videos captured on mobile
devices. Recent research, such as the study conducted by [3], has utilized motion capture
(MOCAP) technology in conjunction with specialized software capable of interpreting
human motion and applying it to biomechanical models. This technology holds significant
potential for enhancing skaters' performance and technique. Additionally, it offers an
effective way to optimize skaters’ training and reduce the risk of injury by focusing on
proper posture, movement form, and execution, similar to ergonomic assessment
approaches used in sports biomechanics studies. However, it is important to note that this
process is time-intensive, as it involves recording multiple video sequences and subsequent
manual review [4].
Identifying specific skating techniques manually from video is expert-dependent and can
be subject to subjectivity, inconsistency, and fatigue, especially with large datasets. These
issues hinder timely, objective feedback from coaches and analysts. Machine learning (ML)
offers a more scalable and consistent alternative by automating movement pattern
recognition and classification for performance evaluation.
In recent years, there has been a growing body of research exploring the advantages of
employing artificial vision techniques and ML in sports [5],[6]. ML enables the analysis of
extensive datasets, extracting patterns and features with minimal human intervention.
Research in the field of sports has harnessed ML techniques such as Support Vector
Machines (SVM) and Convolutional Neural Networks (CNN) to support performance
analysis and movement classification. However, limited research has focused on the
automated classification of skating push techniques in speed skating using pose estimation
and ML approaches.
This research proposes using the OpenPose computer vision system to obtain two-
dimensional coordinates of a skater's joints as they move on the track and to classify the
type of push they use. The main objective is to evaluate the performance of various ML
algorithms to identify and classify three skating push techniques: classic push, double push,
and pendulum push from video sequences. These skating techniques present distinct
biomechanical characteristics associated with propulsion generation, balance control,
movement efficiency, and skating performance. This non-invasive approach allows
measurement of various biomechanical parameters and aims to differentiate a skater's
skating techniques by detecting patterns and trends during motion. To achieve this, tests
and experiments are conducted to assess the system's accuracy and reliability in detecting
and classifying skating techniques. A computer vision and artificial intelligence system for
skating biomechanics analysis enables precise spatio-temporal parameter extraction and the
classification of techniques used by skaters.
In this research, several machine learning techniques are employed for the classification
task. Classification algorithms can identify patterns and features in a training dataset and
Novasinergia 2026, 9(2), 148-173 150
automatically classify new data into the corresponding categories. A Convolutional Neural
Network (CNN) is a neural network architecture composed of multiple convolutional layers
that apply convolutions between filters (or kernels) and the input image. Each kernel is a
matrix whose dimensions depend on the number of channels in the input image and can
extract specific features, such as edges, textures, or shapes [7]. Additionally, there are
pooling layers, which reduce the dimensionality of the output from a layer, and dense
layers, which are fully connected and are used for final classification. CNNs use an
optimization algorithm called Stochastic Gradient Descent (SGD) to adjust the network
weights so that the outputs closely approximate the training data labels [8]. VGG19 is a deep
convolutional neural network widely used in image classification, consisting of 19 layers,
including 16 convolutional layers and three fully connected layers [9].
A Support Vector Machine (SVM) is a supervised learning model used to determine the
class or category to which an input data point belongs. Its objective is to find the hyperplane
that best separates classes in a high-dimensional space, maximizing the margin between
them, thereby classifying new data points based on their position relative to the identified
hyperplane [10]. In classification with non-linear data, a kernel function is used to map the
data to a feature space, enabling their separation by a hyperplane in a higher-dimensional
space [11]. Random Forest (RF) is an ensemble learning method that constructs multiple
decision trees during training and outputs the class that is the mode of the individual trees,
reducing overfitting and improving generalization.
The main contribution of this study is the comparative evaluation of two different
classification paradigms for skating push recognition: a vision-based deep learning
approach using SGEI representations with a CNN based on VGG19, and biomechanical
feature-based machine learning approaches using RF and SVM classifiers. This comparison
provides insights into the advantages, limitations, computational efficiency, and
classification performance of both approaches for sports biomechanics analysis.
2. Methodology
This section describes the methodology used to classify push techniques in speed
skating. The study follows a structured pipeline consisting of four stages: (1) data
acquisition and preprocessing, (2) feature engineering and parameterization, (3) model
selection and training, and (4) model validation and performance evaluation. The
methodology focuses on two classification approaches: image-based classification using
SGEI representations with a CNN, and feature vector-based classification using
biomechanical variables with SVM and RF algorithms.
2.1. Data Acquisition and Preprocessing
Speed skating is a sport that relies on lateral and diagonal gliding on skates, where
mastering an efficient technique is essential to optimize physical capacity [12],[13],[14]. This
research studies the straightway skating technique, which comprises five phases: Basic
Position, Push, Takeoff, Recovery, and Regain. Each phase plays a fundamental role in the
skating technique and contributes to the smoothness of the movement (Figure 1). Three
types of push are identified for classification: (a) the classic push, where the support skate
Novasinergia 2026, 9(2), 148-173 151
remains perpendicular to the ground (Figure 2); (b) the pendulum push, where one skate
pushes while the opposite skate exerts a pull with an inclination of approximately +/-25
degrees (Figure 3); and (c) the double push, characterized by a constant movement in the
supporting leg with a smooth transition between push and pull, shifting the body laterally
with an inclination of approximately +/-25 degrees (Figure 4) [15], [16]. These skating push
techniques exhibit distinct biomechanical characteristics associated with propulsion
generation, balance control, lateral body displacement, and skating efficiency, making them
suitable for automated movement classification and biomechanical performance analysis.
a. Types of Push
The push refers to the force a skater exerts against the track's surface to propel themselves
along their trajectory. The direction of the push may vary depending on the skater's speed
and the slope of the surface [15]. Differences in push execution directly affect skating
propulsion, stability, and movement efficiency, making push classification relevant for
biomechanical analysis and performance optimization. There are three types of push:
In the classic push (Figure 2), the support skate remains perpendicular to the ground.
Figure 2. Classic push (right cycle)
The pendulum push occurs when one skate pushes, while the opposite skate exerts a pull
with an inclination of approximately ±25° (Figure 3).
Figure 3. Pendulum push (right cycle)
During the double push, there is constant movement of the supporting leg, with a smooth
transition between push and pull that shifts the body laterally at an inclination of
approximately ±25 degrees (Figure 4) [16].
Figure 4. Double push (right cycle)
Novasinergia 2026, 9(2), 148-173 152
b. Data Acquisition Protocol
Three mobile cameras are used: two for the sagittal view and one for the frontal view. The
rear camera of an iPhone X, at 60 fps and Ultra High Definition (UHD) resolution, is used
for the front view. Two identical devices, the Xiaomi Redmi Note 10 Pro, at 240 fps and HD
resolution, are employed for the sagittal view. Skateboard videos for this study were
captured using smartphones to simplify data collection, preserve the naturalness of athletes'
movements, and align with the project's budgetary constraints. While professional digital
cameras with advanced subject recognition were initially considered, preliminary testing
revealed unsatisfactory performance in real-world skating conditions. In contrast, modern
smartphones with high-resolution sensors and image stabilization deliver superior video
quality. This practical and cost-effective approach enabled flexible data acquisition without
disrupting athletes, and, despite potential limitations in precision compared to high-end
vision systems, the image quality afforded by smartphones was deemed sufficient for the
goals of this phase.
The video acquisition equipment is mounted on tripods and set to a height of 1.26 meters.
The two side cameras are positioned 6.88 meters from the track and 11 meters apart. The
front camera is placed 9.32 meters from the section delimited by cones where skaters
perform the straight technique. These measurements were defined empirically. The area of
the track where the video is captured is delimited by signaling cones to account for the
athletes' lateral movement.
Figure 5. Distribution of cameras and reference marks.
Background music was used to facilitate subsequent video synchronization using audio in
Filmora 12. The mentioned software enables precise cuts and consistent merging of different
recording angles by using videos that share the same audio.
The participants' ages range from 10 to 24 years, with six male participants and thirteen
female skaters. Each athlete performs one of the three possible push techniques during
straight-line skating.
c. Data Pre-Processing
The collected data is processed to extract information on joint and extremity movement to
classify the type of push a skater uses. This procedure, illustrated in Figure 6, involves
feeding the videos to the VGG19/CNN to get an SGEI graph and a set of feature vectors.
Novasinergia 2026, 9(2), 148-173 153
Figure 6. Process scheme.
d. Video Cutting and Synchronization
The videos collected during the data collection process do not have the same duration.
Furthermore, they capture information both before and after the execution of the straight-
line roller-skating technique. This entails cutting the videos and removing the portions
before and after the roller-skating process in the designated area. The synchronization
process involves temporally aligning two video sequences so they match in time and play
back coherently.
e. Skeleton Extraction
Skeleton extraction is performed using OpenPose, which relies on CNNs to determine its
output information. When a video is input to OpenPose, it generates output in both video
and JSON formats, providing 2D coordinates (x, y) describing the locations of key points.
f. Error Correction
This stage involves correcting errors in joint coordinate data extracted from JSON outputs
generated by OpenPose. Accurate preprocessing at this level ensures the integrity of the
kinematic features derived from the data, which directly influences the downstream
performance and generalization capability of the classification algorithms. Three main types
of errors are identified:
Type 1: Occurs when OpenPose fails to detect a specific keypoint in a frame and
assigns a default value of (0, 0).
Type 2: Results from incorrect or noisy coordinate estimations due to motion blur,
lighting inconsistencies, or shadows.
Type 3: Involves body part swapping, where coordinates are mistakenly assigned to
the opposite side of the body (typically the legs).
Novasinergia 2026, 9(2), 148-173 154
To address these issues, the coordinate data are transformed into a time series. Figure 7
shows two such time series of the x-coordinates of the left and right ankles in sagittal view.
Each time series is analyzed independently.
Figure 7. (a) Time series for the x coordinates of the ankle points, Sagittal View. (b) Separation into error-free intervals
and intervals with errors.
For Type 1 errors, interpolation is used to estimate realistic values. For other error types, we
assess whether a data point deviates significantly from its previous value. Equations (1) and
(2) define dynamic thresholds used to flag such deviations:
thrup = d + 2·sd (1)
thrdo = d − 2·sd (2)
where d is the mean of the differences between consecutive time-series values, and sd is
their standard deviation. A point is flagged as an outlier if the difference from the previous
value exceeds the upper threshold thrup, or falls below the lower threshold thrdo. This
approach, based on a 2-standard-deviation criterion, captures approximately 95% of values
under the assumption of normality and effectively identifies noise and mislabeled data.
To further refine detection, an empirical rule is applied: intervals with fewer than 10
consecutive valid values are classified as error zones, while longer sequences are considered
reliable. Once identified, these erroneous segments undergo an iterative smoothing process,
illustrated in Figure 8, which improves the consistency of the biomechanical features
extracted for subsequent classification.
Figure 8. Flowchart for error correction through iterative smoothing.
Smoothing is performed using a discrete Hanning window of size 10 to preserve local
trends. The affected regions are replaced with the smoothed values. This is repeated
Novasinergia 2026, 9(2), 148-173 155
iteratively, up to a maximum of 20 iterations, or until convergence is achieved. Figure 9
shows the outcome of this process.
Figure 9. Iterative Smoothing for Error Correction in Time Series.
g. Data Smoothing
To obtain a clearer view of the time series trend, a smoothing process is implemented using
a Hanning window. This process involves convolving the time sequence with the window,
allowing a more precise representation of the series' behavior.
h. Data Normalization
The smoothed data is normalized to ensure they all fall within a common scale. This process
eliminates magnitude disparities among the data, thus facilitating their comprehension and
analysis (Figure 10).
Figure 10. (a) Data from an individual in two different frames without normalization. (b) Data from an individual in two
different frames after applying normalization.
Novasinergia 2026, 9(2), 148-173 156
2.2. Feature Engineering and Parameterization
a. OpenPose and the Tools for Motion Capture
The MOCAP system enables the transfer of human motion to biomechanical models, both
two-dimensional and three-dimensional, through a combination of specialized hardware
and software [3].
On the other hand, there is OpenPose, a computer vision system capable of real-time
detection and tracking of human bodies through RGB cameras and processing previously
recorded images and videos. OpenPose relies on two main components: keypoint
confidence maps and Part Affinity Fields (PAFs). Confidence maps indicate the probability
of the presence of a keypoint, while PAFs are vector fields that contain information about
the orientation and direction between two keypoints [17].
b. Pose Data Normalization
When capturing a person's movement on video, their relative size and position within the
frame may vary due to differences in body proportions and camera angles. To ensure that
the analysis remains consistent across individuals and recording conditions, a
normalization step is applied to the skeletal joint coordinates.
Following the approach proposed in [18], [19], Equation (3) is used to normalize each joint
position by centering it around the hip and scaling it with respect to the body height. This
allows us to characterize the skeleton independently of its absolute location or size in the
image, making it suitable for comparison and classification.
j′t,i = (xt,i − x cm t , yt,i − y cm t ) / ht(jN t , jtm t ) (3)
where t denotes the frame number, j′t,i = (xt,i, yt,i) (i = 0,…,24) represents the normalized
coordinates of joint i, cm refers to the hip center, and ht is the Euclidean distance from the
nose (jN t) to the midpoint between the left and right ankles (jtm t).
This normalization process is essential to our objective of classifying skating push
techniques, as it ensures that joint movement data are expressed on a common scale and are
thus directly comparable across different skaters. Reducing inter-subject variability
improves the accuracy and generalizability of the machine learning models used in
subsequent stages.
c. Gait Energy Image (GEI)
The GEI is a widely used representation that summarizes an individual's movement over a
gait cycle by averaging binary silhouettes across time. It effectively captures the overall
posture and motion dynamics of the subject, making it valuable for tasks such as person
identification and abnormal gait classification.
In our study, the GEI is used to create a structured, low-dimensional visual representation
of the skater's movement that serves as input for the CNN-based classification system
described in Section Results. By condensing the entire movement into a single image, the
Novasinergia 2026, 9(2), 148-173 157
GEI enables the convolutional neural network to learn spatial patterns and distinctive traits
associated with different types of push techniques.
The GEI is computed following the formulation proposed in [20], as shown in Equation (4):
GEI(x, y) = (1/N) Σt=1→N Bt(x, y) (4)
where Bt(x, y) represents the binary pixel value at coordinates (x, y) in frame t, and N is the
total number of frames in a gait cycle.
A variant called the Skeleton Gait Energy Image (SGEI) is also employed in this work.
Unlike the traditional GEI, which uses full-body silhouettes, the SGEI is generated from
skeleton silhouettes obtained using OpenPose [21]. This alternative focuses on joint
movement and posture, offering a more abstract and flexible representation, particularly
well-suited to sports movement classification.
d. Extraction of Biomechanical Features
Extracting biomechanical features involves identifying and quantifying athletes' movement
patterns, considering joint angles, velocities, and distances. Two approaches are used to
extract biomechanical features: the graphical approach, which uses SGEI images for each
skater classified by a CNN, and the data-driven approach, which uses feature vectors and
implements a classifier with an SVM.
e. Obtaining the SGEI
Images are generated using SGEI from binary silhouette skeletons obtained for each frame
of every video and corresponding person. The SGEI combines and summarizes the athlete's
information and movement into a single image, enabling visualization of the distinctive
characteristics of each thrust type, as shown in Figure 11. This technique provides a clear,
consolidated view of the athlete's movements, effectively highlighting the unique features
of different thrust types.
Figure 11. (a) Front view. (b) Sagittal view. SGEI images generated from binary silhouettes.
Novasinergia 2026, 9(2), 148-173 158
f. Calculation of Spatio-Temporal Variables and Angles
During the execution of a straight-line thrust, changes and interactions among relevant
kinematic variables occur, providing valuable insights. The complete execution,
encompassing both the right and left periods, is considered a full cycle. Variables such as
distances, times, velocities, and amplitudes are analyzed. Additionally, Table 1 and Figure
12 show the angles corresponding to the frontal and sagittal views. This approach offers a
comprehensive examination of motion dynamics, shedding light on the intricate
relationships between these kinematic variables.
Figure 12. (a) Front view. (b) Sagittal view. Calculated angles.
Table 1. Angles
IDENTIFIER
SAGITTAL
FRONTAL
a
Right knee
Right outside foot
b
Left knee
Left outer foot
c
Right shin
Right inner foot
d
Left shin
Left inner foot
e
Right hip
Right Inner Shoulder Push
f
Left hip
Left Inner Shoulder Push
g
Right shoulder
Right ankle push
h
Left shoulder
Left ankle push
i
Trunk
Right external shoulder push
j
Right ankle
Left external shoulder push
k
Left ankle
Crotch
l
Crotch
–
m
Right elbow
–
n
Left elbow
–
2.3. Model Selection and Training
For the CNN-based approach, a classification model was built using the VGG19
architecture as the foundation and implemented in Python with TensorFlow Keras. Transfer
learning was utilized by retaining the pre-trained weights in the initial layers, which detect
common patterns, while the subsequent layers adjust their weights based on the skating
dataset. Given the limited size of the database, data augmentation techniques were applied
Novasinergia 2026, 9(2), 148-173 159
to the input images, including rotation, shifting, and zooming. A Dropout regularization
layer was included by randomly deactivating a percentage of neurons during each training
epoch to prevent overfitting and improve generalization. The specific hyperparameters of
the CNN model are as follows: batch size of 32 to minimize excessive memory usage, SGD
optimizer with a learning rate of 0.002, and categorical cross-entropy as the loss function.
The VGG19 base model processes 224x224 pixel images through 16 convolutional layers
using multiple 3x3 filters, followed by three fully connected layers for final classification.
Figure 13. Architecture of the proposed VGG19-based transfer learning model for skating push classification using SGEI
representations.
To implement transfer learning, the convolutional base of VGG19 pre-trained on ImageNet
was used as a feature extractor. The initial convolutional blocks were frozen to preserve
previously learned low-level visual features such as edges, textures, and shapes. The final
convolutional block was fine-tuned using the skating dataset to adapt the model to the
specific movement patterns present in SGEI images. After the convolutional base, a Global
Average Pooling layer, a dense layer with ReLU activation, and a Dropout layer with a rate
of 0.5 were added to reduce overfitting and improve generalization. Finally, a Softmax
output layer with three neurons was used to classify the skating push techniques: classic
push, double push, and pendulum push.
For the SVM-based approach, the feature vectors comprising 95 biomechanical variables
were classified using the scikit-learn (sklearn) library. The SVM was configured with a
Radial Basis Function (RBF) kernel, selected due to the non-linear nature of the data. The
regularization parameter C was set to 0.05, allowing for a wider margin and preventing
overfitting, and the gamma parameter was set to 0.004, controlling the influence of each
training vector on the construction of the decision hyperplane. These hyperparameter
values were selected through iterative search to optimize the balance between model
complexity and generalization capability.
For the Random Forest classifier, the same feature vector set was used. The key
hyperparameters were configured as follows: n_estimators was set to 100, providing a
reasonable balance between computational performance and model precision; the splitting
criterion was the Gini impurity index; and a random_state of 1 was set to ensure
Novasinergia 2026, 9(2), 148-173 160
reproducibility of the results. No maximum depth constraint was applied, allowing trees to
grow until all leaves were pure or contained fewer than the minimum samples required for
splitting.
2.4. Model Validation and Performance Metrics
To evaluate the classification performance of the three models, the dataset was
partitioned into training, validation, and testing subsets. For the CNN approach, a total of
2,618 SGEI images were divided as follows: 70% for training, 15% for validation, and 15%
for testing. This distribution ensures a comprehensive learning and evaluation process,
allowing the model to learn from the training set, tune hyperparameters using the validation
set, and assess generalization on unseen test data.
For the SVM and RF approaches, a total of 264 feature vectors were processed, with 75%
used for training and 25% for testing. To address class imbalance among the three push
types (64 classic, 77 double, 123 pendulum), a synthetic data generation process (SMOTE-
like oversampling) was applied to the training set, resulting in 240 balanced training vectors
per class.
The following performance metrics were used to evaluate all classifiers: (a) overall accuracy,
defined as the proportion of correctly classified instances over the total number of instances;
(b) confusion matrix, providing a detailed breakdown of true positives, false positives, true
negatives, and false negatives for each class; and (c) Receiver Operating Characteristic
(ROC) curves with Area Under the Curve (AUC) values, which assess each classifier
discriminative ability across varying decision thresholds. Additionally, training accuracy
was reported alongside test accuracy to assess potential overfitting, and execution time was
measured to compare computational efficiency across classifiers.
3. Results
This section presents the results obtained from the data processing and classification
pipeline described in the Methodology. The results are organized in three parts: first, the
biomechanical analysis including spatio-temporal parameters and angular measurements;
second, the comparative analysis of lower extremity angles across push types; and third, the
classification performance evaluation for each machine learning model, including the
optimal configurations achieved and the performance metrics obtained on the test sets.
3.1. Biomechanical Analysis
Tables 2, 3, and 4 present the mean and standard deviation (±) of spatio-temporal
parameters and angles. The label (D) refers to the right side of the body and (I) to the left
side. The results demonstrate a notable similarity to previous skating research conducted
by [3] and [16], validating the data obtained through the developed system. Table 2 reveals
that the speed of double push (9.34 m/s) exceeds that of the classic push (8.26 m/s) and
pendulum push (8.62 m/s). When adjusting distances to a common scale for ease of
comparison, time and speed values for different pushes are obtained: classic (1.82 s and 8.06
m/s), double (1.61 s and 9.12 m/s), and pendulum (1.75 s and 8.38 m/s). This confirms that
the double push continues to have the shortest time. The double push is more efficient
Novasinergia 2026, 9(2), 148-173 161
compared to other types of push, as it achieves higher speed with less effort. This is due to
the simultaneous use of both feet and improved movement synchronization, resulting in
greater propulsion and speed, indicating more efficient energy use and overall
performance.
Table 2. Distances, times, and speeds (average ± standard deviation)
CHARACTERISTICS
DOUBLE
PENDULUM
Max Lateral Amplitude D (m)
0.55 ± 0.12
0.55 ± 0.13
Max Lateral Amplitude I (m)
0.45 ± 0.11
0.44 ± 0.10
Half-Cycle Time D (s)
0.71 ± 0.22
0.81 ± 0.29
Half-Cycle Time I (s)
0.83 ± 0.22
0.90 ± 0.31
Full Cycle Time (s)
1.60 ± 0.41
1.75 ± 0.51
Half-Cycle Distance D (m)
6.50 ± 1.65
6.76 ± 1.97
Half-Cycle Distance I (m)
7.59 ± 1.62
7.51 ± 2.13
Cycle Distance (m)
14.57 ± 2.55
14.68 ± 3.27
Speed (m/s)
9.34 ± 1.26
8.62 ± 1.09
Max shoulder-to-knee distance D (m)
0.63 ± 0.08
0.69 ± 0.07
Max shoulder-to-knee distance I (m)
0.72 ± 0.06
0.72 ± 0.06
Min shoulder-to-knee distance D (m)
0.36 ± 0.04
0.39 ± 0.04
Min shoulder-to-knee distance I (m)
0.37 ± 0.05
0.37 ± 0.05
Max Frontal Amplitude D (m)
0.50 ± 0.09
0.49 ± 0.09
Max Frontal Amplitude I (m)
0.48 ± 0.10
0.41 ± 0.07
Min Frontal Amplitude D (m)
0.11 ± 0.05
0.17 ± 0.05
Min Frontal Amplitude I (m)
0.13 ± 0.05
0.17 ± 0.07
Mid-shoulder distance (m)
0.27 ± 0.02
0.26 ± 0.02
Max foot-to-ground clearance D (m)
0.19 ± 0.05
0.25 ± 0.08
Max foot-to-ground clearance I (m)
0.18 ± 0.05
0.26 ± 0.07
Table 3. Mean values of the angles in front view (average ± standard deviation)
CHARACTERISTICS
CLASSIC (°)
DOUBLE (°)
PENDULUM (°)
Outer foot D
84.71 ± 6.53
89.04 ± 4.70
88.91 ± 4.91
Outer foot I
79.56 ± 4.53
82.82 ± 4.63
79.81 ± 5.25
Inner foot D
95.29 ± 6.53
90.96 ± 4.70
91.09 ± 4.91
Inner foot I
100.44 ± 4.53
97.18 ± 4.63
100.19 ± 5.25
Inner Push D
10.71 ± 1.86
12.02 ± 2.03
12.92 ± 2.17
Inner Push I
7.80 ± 1.86
8.99 ± 1.92
9.57 ± 2.40
Outer Push D
18.72 ± 2.47
19.20 ± 1.74
20.31 ± 3.17
Outer Push I
17.03 ± 2.50
19.71 ± 1.62
19.49 ± 2.69
Ankle D
80.36 ± 4.27
82.04 ± 3.32
78.88 ± 4.79
Ankle I
81.51 ± 4.22
84.15 ± 3.11
81.74 ± 4.40
Groin
47.57 ± 6.24
46.58 ± 5.88
48.49 ± 5.85
Novasinergia 2026, 9(2), 148-173 162
In Table 3, the average angles of the outer foot for the right (D) (89.04°) and left (I) (82.82°)
indicate a greater inclination in the double push, with a more pronounced angle than other
push types. The angles for internal (D at 12.92° and I at 9.57°) and external (D at 20.31° and
I at 19.49°) pushes illustrate the distinct, non-simultaneous traction and push characteristics
of the pendulum push. Table 4 shows that double push involves simultaneous traction and
push, thereby reducing the need for excessive knee bending, with average knee angles of
right (D) (118.66°) and left (I) (120.3°). However, this type of push requires greater trunk
flexion (17.87°) than the classic and pendulum pushes.
Table 4. Mean values of the angles in sagittal view (average ± standard deviation)
CHARACTERISTICS
CLASSIC (°)
DOUBLE (°)
PENDULUM (°)
Trunk
21.25 ± 4.19
17.87 ± 2.74
23.45 ± 4.64
Hip D
86.63 ± 4.52
80.13 ± 3.93
87.42 ± 6.30
Hip I
83.65 ± 6.24
81.15 ± 5.45
85.09 ± 8.73
Knee D
119.01 ± 7.10
118.66 ± 4.42
116.19 ± 7.31
Knee I
118.18 ± 5.71
120.30 ± 5.72
116.33 ± 6.53
Ankle D
102.23 ± 8.62
101.20 ± 4.05
99.88 ± 5.84
Ankle I
97.58 ± 5.06
101.10 ± 4.59
98.35 ± 6.41
Shin D
53.80 ± 6.90
57.74 ± 3.13
52.25 ± 4.99
Shin I
54.85 ± 5.90
55.65 ± 6.09
54.14 ± 6.78
Groin
31.29 ± 7.46
31.18 ± 6.91
33.10 ± 7.34
Elbow D
125.94 ± 18.84
132.01 ± 17.05
116.27 ± 25.44
Elbow I
141.96 ± 14.82
129.60 ± 21.50
141.20 ± 14.75
Shoulder D
51.57 ± 12.15
41.24 ± 15.23
53.07 ± 17.62
Shoulder I
33.83 ± 10.14
27.80 ± 12.82
25.17 ± 17.43
These findings reveal bilateral concordance between the right and left half-cycles, as values
on the right side of the body do not differ significantly from those on the left. This indicates
a suitable biomechanical balance in the execution of the analyzed movements, as well as an
equitable distribution of weight and movement across the lower and upper limbs on both
sides of the body. However, it is important to note that results can vary significantly based
on the participants' height, weight, flexibility, strength, and coordination. Each individual
is unique, and what benefits one person may not necessarily apply to another, depending
on their personal characteristics. This observation underscores the importance of
considering individual differences when analyzing movement efficiency and effectiveness.
3.2. Comparison of the Angles of the Lower Extremities
Graphs showing the angle as a function of the speed skating cycle are generated,
based on the angles of the trunk flexion, knee, and ankle in the sagittal view. The colored
regions illustrate the angle variations of the 19 volunteer athletes, while the dark lines
represent the average. The red graphs correspond to the classic push, the green graphs
denote the double push, and finally, the blue graphs represent the pendulum push. These
visual representations provide a clear, detailed comparison of the biomechanical dynamics
Novasinergia 2026, 9(2), 148-173 163
of each push technique, highlighting the distinct movement patterns athletes use in speed
skating.
The patterns among the different types of pushes are quite similar; however, there are
differences in the flexion range at various parts of the cycle. Figure 14 illustrates that during
the recovery phase, which spans from 40% to 80% of the cycle, there is a greater trunk
flexion. This increased flexion is due to the weight transfer during the push phase. While in
the classic push, the recovery phase occupies 42% of the cycle, the pendulum push exhibits
less-defined values and a wider range of trunk movement.
Figure 14. Trunk flexion: (a) Classic push. (b) Double push. (c) Pendulum push.
Figure 15 shows the patterns in the angles of the knee joint of the left leg. Flexion is observed
during the push phase, which occurs at 20% of the cycle, meaning a decrease in the angle.
At this point, the majority of body weight is supported by the left leg, resulting in maximum
flexion during the recovery phase, which occurs from 25% to 80% of the cycle. The final 20%
corresponds to the final phase, in which the leg returns to its initial support state. It is noted
that in the classic and pendulum pushes, there is greater knee flexion compared to the
double push.
Figure 15. Left knee flexion: (a) Classic push. (b) Double push. (c) Pendulum push.
Figure 16 illustrates the pattern formed by the angle of the right ankle. A more defined
pattern is observed in the double push, characterized by a more consistent standard
deviation, attributable to its steady incline movement. On the contrary, the classic push
exhibits a flattened flexion, reaching a maximum of 125 degrees. In contrast, the pendulum
push exhibits a pattern that combines features of both the classic and double pushes.
Novasinergia 2026, 9(2), 148-173 164
Figure 16. Right ankle flexion: (a) Classic push. (b) Double push. (c) Pendulum push.
3.3. Classification Performance Evaluation
This section presents the classification results for each of the three machine learning
approaches evaluated: CNN, SVM, and RF. Each subsection reports the data allocation
strategy, optimal hyperparameter configurations, and final performance metrics achieved
on the respective test sets. The three push types are numerically identified as classic (0),
double (1), and pendulum (2).
a. Classification Using CNN
A total of 2618 Silhouette-based Gait Energy Images (SGEI) were used as input for
classification, which was conducted manually by an expert. These images were categorized
into 759 classic push, 963 double push, and 896 pendulum push images. For each push type,
70% of the images were assigned to training, 15% to validation, and the remaining 15% to
testing. This distribution ensures a comprehensive learning and evaluation process,
allowing the models to effectively learn from and adapt to the varied characteristics of each
push type.
A batch size of 32 was chosen to minimize excessive memory usage. The model employs an
SGD optimizer with a learning rate of 0.002. Additionally, the categorical cross-entropy loss
function was used. The CNN-based classification achieved 80.97% accuracy on the training
set and 88.92% on the validation set. For testing, 377 Silhouette-based Gait Energy Images
(SGEI) were used, including 91 classic push images, 143 double push, and 143 pendulum
push images. The results of the confusion matrix and Receiver Operating Characteristic
(ROC) curves are presented in Figure 17, which showcase an accuracy of 90.72% and an area
under the curve of 98.24%.
Novasinergia 2026, 9(2), 148-173 165
Figure 17. CNN evaluation: (a) Confusion matrix. (b) ROC curve.
b. Classification Using SVM
The feature vector set is classified using the SVM method with an RBF kernel, selected due
to the non-linear nature of the data. The C parameter is set to 0.05, allowing for a wider
margin, and the Gamma parameter is adjusted to 0.004, thus increasing the influence of each
training vector on the construction of the decision hyperplane. This configuration is
carefully chosen to balance the model's complexity with the specificity required to
accurately classify the non-linear dataset, thereby optimizing the SVM classifier's
performance.
A total of 264 feature vectors were processed: 64 corresponding to the classic push class, 77
to the double push class, and 123 to the pendulum push class. These data were divided into
a training set comprising 75% of the total and a testing set comprising the remaining 25%.
A synthetic data generation process was applied, yielding 240 training vectors per class. An
accuracy of 92.22% was achieved on the training data and 93.94% on the test set.
Figure 18. SVM evaluation: (a) Confusion matrix. (b) ROC curve.
Novasinergia 2026, 9(2), 148-173 166
c. Classification Using Random Forest
Using the same training and testing sets as for the Support Vector Machine (SVM), an
accuracy of 100% was achieved on the training set and 92.42% on the testing set, suggesting
potential overfitting of the classifier. Figure 19(b) reveals that the curve for the pendulum
class is not as convex as that for the other two classes, which may indicate problems with
class imbalance. This observation points to the classifier's high proficiency on the training
data, while also highlighting the need for caution regarding its generalization to unseen
data, particularly in effectively managing class imbalance.
Figure 19. Random Forest evaluation: (a) Confusion matrix. (b) ROC curve.
4. Discussion
The three classification systems presented in this study demonstrate high accuracy
on their respective test sets: 90.72% for CNN, 93.94% for SVM, and 92.42% for RF. The
confusion matrices and ROC curves indicate promising results for classifying skater push
types during the straight-line technique. These results position our work competitively
within the growing body of literature on machine learning applications for sports
movement analysis and biomechanical classification.
Our SVM-based approach, achieving 93.94% test accuracy with an RBF kernel (C=0.05,
gamma=0.004), is consistent with findings reported in related domains. For instance, the
integration of OpenPose with SVM for postural analysis in young adults and achieved high
classification accuracy using temporal and spatial regression features [22]. Similarly, in
combat sports, SVM-based classification of Taekwondo kick types using accelerometer data
archieved accuracies exceeding 96% [23]. While our accuracy is slightly lower, this is likely
attributable to the higher complexity of skating push classification, which involves subtle
differences in body inclination and timing rather than distinctly different movement
patterns. The use of smartphone-based video capture, while offering practical advantages
in accessibility and cost, introduces additional noise compared to dedicated motion capture
systems or wearable inertial sensors.
The CNN approach, using VGG19 transfer learning with SGEI inputs achieved 90.72%
accuracy, which aligning with results from similar image-based classification systems in
sports. A comprehensive review of pose estimation models and found that OpenPose-based
Novasinergia 2026, 9(2), 148-173 167
approaches in sports applications generally achieve accuracy levels above 85% when
combined with appropriate deep learning classifiers [24]. Similary, the application of
OpenPose for pose estimation in hurdles athletics demonstrated sufficient accuracy for
practical athlete development applications [25]. The relatively lower CNN accuracy
compared to SVM may be explained by the limited dataset size (2,618 images), as deep
learning models typically require larger datasets to fully exploit their capacity for learning
complex visual features.
The RF classifier achieved 100% training accuracy but 92.42% test accuracy, suggesting
potential overfitting despite its overall strong performance. This pattern is consistent with
observations from comparative studies of SVM, RF, KNN, and Logistic Regression for gait
classification, which found that ensemble methods tend to overfit when applied to relatively
small datasets [26]. The class imbalance in our feature vector dataset (64 classic, 77 double,
123 pendulum vectors) may have contributed to the RF overfitting tendency, even after
applying synthetic data generation. Future work should explore more sophisticated class
balancing strategies, such as SMOTE with Tomek links or cost-sensitive learning
approaches.
The biomechanical analysis results are consistent with prior skating research [3], [16],
thereby validating the data obtained through the developed system. The finding that double
push achieves a higher speed (9.34 m/s) than classic push (8.26 m/s) and pendulum push
(8.62 m/s) aligns with established biomechanical principles, as the double push leverages
the simultaneous use of both feet for greater propulsion efficiency. The higher classification
performance observed for the “double push” technique may be associated with its more
distinctive biomechanical movement pattern, characterized by continuous lateral
displacement and smoother transitions between push and pull phases. In contrast, the
“classic push” and “pendulum push” present more similar movement characteristics, which
may increase classification complexity. Two-dimensional video-based pose estimation has
been shown to reliably capture gait parameters comparable to marker-based systems [27],
supporting the validity of our OpenPose-based approach for extracting meaningful
biomechanical features from video data.
The use of smartphone cameras for data acquisition represents a significant practical
advantage of our approach. A similar framework that utilizes smartphone monocular
videos for gait analysis and demonstrated that this cost-effective alternative can provide
reliable kinematic measurements [28]. Furthermore, the feasibility of using OpenPose for
markerless motion analysis in real athletics competitions han been demonstrated,
confirming that markerless pose estimation from standard video can be practically
deployed in field settings [29]. Our results extend these findings to the specific domain of
speed skating technique classification.
The feature vector approach with 95 biomechanical variables outperformed the image-
based CNN approach, suggesting that domain-specific feature engineering retains
significant value even in the era of deep learning. Among the evaluated models, the SVM
classifier achieved the best overall performance, obtaining the highest test accuracy (93.94%)
and the shortest execution time, suggesting that biomechanical feature-based approaches
are highly effective for skating push classification. This observation is supported by findings
indicating that bridging the lab-to-field gap in sports biomechanics often requires careful
Novasinergia 2026, 9(2), 148-173 168
integration of domain knowledge with machine learning techniques [30]. Similarly,
wearable biomechanical analytics combined with machine learning can enhance sports
performance monitoring when appropriate feature selection is applied [31]. The 95-variable
feature set used in our study captures joint angles, velocities, and spatio-temporal
parameters that directly encode the biomechanical differences between push types.
When comparing computational efficiency, SVM demonstrated significantly shorter
execution time (0.692 s) compared to CNN (173.689 s) and RF (2.349 s). This is consistent
with the computational characteristics of these algorithms: SVMs with kernel functions are
generally efficient for moderately dimensional feature vectors, whereas CNNs require
substantial computational resources for image processing during training. The practical
implications are important for potential real-time deployment; the SVM classifier combined
with a pre-trained OpenPose model could feasibly be deployed for near-real-time feedback
during training sessions, as suggested by recent work on edge-computing wearables for
sports analytics [32],[33].
Several studies have applied deep learning to related sports analysis tasks with comparable
or higher accuracy. The effectiveness of Gait Energy Images for human gait analysis has
been demonstrated [20], and skeleton-based GEI representations have been shown to
achieve strong classification performance for pathological gait detection [21]. Our
adaptation of SGEI for skating push classification extends these approaches to a novel
application domain. Additionally, robust classification of baseball player behavior using
LSTM networks with multimodal features has been achieved [18], suggesting that temporal
sequence models could improve upon our results by capturing the dynamic evolution of
skating movements across frames.
Recent reviews have highlighted the growing importance of AI in sports biomechanics. A
scoping review spanning 73 studies (2015-2024) found that methodological sophistication
has evolved from traditional ML to deep learning architectures that integrate multimodal
data [34]. A systematic review of ML for sport-specific movement recognition identified that
SVM and RF remain among the most effective traditional approaches for classification tasks
with engineered features [35]. Our study contributes to this body of work by demonstrating
that these traditional approaches, when combined with computer vision-based feature
extraction via OpenPose, can achieve competitive results in a previously unexplored
application domain.
The present study has several limitations that should be acknowledged. First, the dataset
comprises 19 participants aged 10-24 years, which, while sufficient for the classification task
demonstrated, limits the statistical power and generalizability of the findings. Second, using
2D pose estimation from monocular video introduces inherent depth ambiguity,
particularly for out-of-plane movements. Accuracy limitations of markerless motion
capture for measuring running kinematics have been documented [36], and similar
limitations may apply to skating analysis. Third, the hyperparameter optimization was
performed through iterative manual search rather than systematic grid search or Bayesian
optimization, which may have resulted in suboptimal configurations. Fourth, the absence
of k-fold cross-validation limits the robustness assessment of the reported accuracy metrics.
Despite these limitations, this study offers several novel contributions. First, it represents
one of the earliest applications of combined computer vision and machine learning to
Novasinergia 2026, 9(2), 148-173 169
classify push techniques in speed skating, addressing a gap identified in the sports
biomechanics literature. Second, it demonstrates that a non-invasive, smartphone-based
approach can yield reliable biomechanical parameters and accurate technique classification,
making advanced biomechanical analysis accessible to coaches and athletes without
expensive equipment. Third, the comparison of image-based (CNN) and feature-based
(SVM, RF) approaches provides practical guidance for selecting classification strategies
based on available data and computational resources. In particular, the comparative
analysis between the vision-based CNN approach and the biomechanical feature-based ML
approaches highlights differences in classification performance, computational efficiency,
and feature-extraction capabilities for sports biomechanical analysis. Future work should
focus on expanding the dataset, implementing systematic hyperparameter optimization
with cross-validation, exploring temporal sequence models such as LSTM or Transformer
architectures, and validating the system in real-time training scenarios [37],[38],[39].
5. Conclusions
This study demonstrates the feasibility of extracting and processing video data to
obtain Silhouette-based Gait Energy Images (SBGEI) and feature vectors. These elements
have been used to develop training, validation, and testing datasets for machine learning
(ML)-based classification methods of push types in speed skating. The performance metrics
achieved are consistent with findings from similar research efforts demonstrating the
effectiveness of these methods for classifying speed skating techniques. This approach not
only underscores the utility of video analysis in sports science but also highlights the
potential of ML applications to improve our understanding and analysis of athletic
performance.
A study has been conducted on biomechanical models of the straight-line skating technique.
The distinctive characteristics of the classic push, double push, and pendulum push in
skating have been identified, and the main phases of movement during their execution have
been described. This investigation not only elucidates the intricate mechanics of skating
techniques but also provides a foundational understanding to optimize performance
through biomechanical analysis.
OpenPose offers a non-invasive approach for generating dynamic variables relevant to the
study, allowing athletes to perform their techniques without relying on devices that restrict
their mobility. In addition, skateboard videos were captured using smartphones,
eliminating the need for complex or expensive equipment. This method not only facilitates
data collection but also ensures that athletes can maintain natural movement patterns,
providing authentic insights into athletic performance.
In this study, two classification approaches were explored: SGEI using a CNN, achieving an
accuracy of 90.72% with the VGG19 architecture as the foundation, and the feature vector
approach with 95 variables derived from a biomechanical analysis using an SVM, achieving
an accuracy of 93.94%, and the RF algorithm, achieving an accuracy of 92.42%. These
methods demonstrate the potential to combine advanced image processing with
biomechanical insights to precisely classify athletic movements.
Our study innovatively uses everyday smartphone video and machine learning to
effectively classify speed skating push techniques through Silhouette-based Gait Energy
Novasinergia 2026, 9(2), 148-173 170
Images and movement data. Achieving high accuracy with common algorithms confirms
the power of these methods for performance analysis. Crucially, our biomechanical
modeling reveals key, practical differences between push styles using accessible tools like
smartphones and OpenPose, without disturbing athletes. This offers new, real-world
insights for optimizing training and understanding technique in speed skating.
Future work will focus on expanding the dataset to include a greater variety of athletes and
skating conditions to improve the models’ generalizability. Additionally, future studies
could explore the classification of other technical maneuvers, such as cornering or starts, to
provide a more comprehensive analysis of speed skating performance.
Author Contributions
Conceptualization, X.A.-T. and F.A.-S.; methodology, X.A.-T. and S.U.-M.; software,
X.A.-T. and S.U.-M.; validation, X.A.-T., S.U.-M., J.B., L.I.M., and F.A.-S.; formal analysis,
X.A.-T.; investigation, X.A.-T. and S.U.-M.; resources, F.A.-S.; data curation, X.A.-T.;
writing—original draft preparation, X.A.-T. and S.U.-M.; writing—review and editing, J.B.,
L.I.M., and F.A.-S.; visualization, X.A.-T.; supervision, F.A.-S.; project administration, F.A.-
S. All authors have read and approved the published version of the manuscript.
Acknowledgments
We would like to express our sincere gratitude to the Gabriela Cavalieri Club for their
invaluable collaboration in making this work possible. Their commitment and enthusiasm
were crucial for the success of this project. Special thanks go to the talented athletes of the
club, whose stories and experiences greatly enriched our work.
The Club's support extended beyond the participation of its athletes. We also want to
highlight the assistance received from the coaches, parents, and club members, who were
always available to help with anything needed.
Furthermore, we extend our heartfelt appreciation to the Ecuadorian Corporation for the
Development of Research and Academia (CEDIA) for granting us access to its High-
Performance Computing (HPC) cluster. Their support has been instrumental in the
development and completion of this project, enabling us to perform complex calculations
and large-scale data analysis efficiently. Access to CEDIA's HPC has provided an advanced
technological infrastructure indispensable to our research. We also appreciate the technical
team at CEDIA for their continuous assistance and support, which ensures optimal
utilization of these resources. This project would not have been possible without the
collaboration and generosity of CEDIA, and we highly value their commitment to advance
research and academia in Ecuador.
Conflict of Interest
The authors report no conflicts of interest related to this research.
Novasinergia 2026, 9(2), 148-173 171
Generative Artificial Intelligence (AI) Use Statement
No generative artificial intelligence was used in the preparation of this article.
Funding Sources
This research was funded by the University of Cuenca and CEDIA (Ecuadorian
Corporation for the Development of Research and Academia).
References
[1] A. J. Leemreize, “Assessment of performance in speed skating using IMUs,” Master's thesis, Univ. of
Twente, Enschede, Netherlands, 2024. [Online]. Available:
https://essay.utwente.nl/fileshare/file/104719/Leemreize_MA_TNW.pdf
[2] P. Madeleine, A. Samani, M. De Zee, and U. Kersting, “Biomechanics of human movement,” IFMBE
Proc., vol. 34, pp. 237–240, 2011, doi: 10.1007/978-3-642-21683-1_60.
[3] J. Acero, “MÉTODO MOCAP SPORTECH- II&SB PARA LA EVALUACIÓN BIOMECÁNICA DE LA
TÉCNICA DEPORTIVA,” in Mem. X Encuentro Internacional en Ciencias de la Actividad Física y del
Deporte, 2026, pp. 9–10. [Online]. Available:
https://www.unipamplona.edu.co/unipamplona/portalIG/home_145/recursos/general/14042026/xenc
uentropamplona2025.pdf
[4] I. Takeda, A. Yamada, and H. Onodera, “Artificial Intelligence-Assisted motion capture for medical
applications: a comparative study between markerless and passive marker motion capture,” Comput.
Methods Biomech. Biomed. Eng., vol. 24, no. 8, pp. 864–873, 2021, doi: 10.1080/10255842.2020.1856372.
[5] Y. Yao, J. Wang, Y. Lai, and L. C. Chen, “Aging Decline in Basketball Career Trend Prediction Based
on Machine Learning and LSTM Model,” arXiv preprint arXiv:2509.25858, 2025, doi:
10.48550/arXiv.2509.25858.
[6] Z. Lu and X. Mao, “A gait recognition method based on deep learning and attention transformer,” in
Fourth International Conference on Signal Processing and Computer Science (SPCS 2023), vol. 12970, 2023,
doi: 10.1117/12.3012246.
[7] H. H. Aghdam and E. J. Heravi, Guide to convolutional neural networks, Springer International
Publishing, 2017. doi: 10.1007/978-3-319-57550-6.
[8] R. Venkatesan and B. Li, Convolutional neural networks in visual computing: A concise guide. Boca Raton,
FL, USA: CRC Press, 2017.
[9] M. Bansal, M. Kumar, M. Sachdeva, and A. Mittal, “Transfer learning for image classification using
VGG19: Caltech-101 image data set,” J. Ambient Intell. Humaniz. Comput., vol. 14, no. 4, pp. 3609-3620,
sep. 2021, doi: 10.1007/s12652-021-03488-.
[10] J. A. Carter, A. R. Rivadulla, and E. Preatoni, “A support vector machine algorithm can successfully
classify running ability when trained with wearable sensor data from anatomical locations typical of
consumer technology,” Sport. Biomech., vol. 23, no. 11, pp. 2372–2389, jan. 2022, doi:
10.1080/14763141.2022.2027509.
[11] D. A. Pisner and D. M. Schnyer, “Support vector machine,” in Machine Learning, A. Mechelli and S.
Vieira, Eds. Oxford, UK: Academic Press, 2020, pp. 101–121, doi: 10.1016/B978-0-12-815739-8.00006-7.
[12] S. Sehgal et al., “Unleashing potential and optimizing adolescent roller skating performance through
a structured exercise program – a randomized controlled trial,” BMC Sports Sci. Med. Rehabil., vol. 15,
no. 1, sep. 2023, doi: 10.1186/s13102-023-00728-x.
[13] G. J. van Ingen Schenau, R. W. De Boer, and G. De Groot, “Biomechanics of speed skating,” in
Biomechanics of Sport, C. L. Vaughan, Ed. Boca Raton, FL, USA: CRC Press, 2020, pp. 121–167, doi:
10.4324/9781003068549-4
[14] G. Bongiorno, F. G. Minisini, H. Biancuzzi, F. Dal Mas, and L. Miceli, “Skating efficiency and technique
during roller speed skate using innovative piezoelectric smart socks : an exploratory study,” Front.
Sports Act. Living, vol. 7, pp. 1–10, jul. 2025, doi: 10.3389/fspor.2025.1554264.
[15] Z. Liu, M. Ding, M. Zhang, B. Yu, and H. Liu, “Effects of Technique Asymmetry on 500 m Speed
Skating Performance,” Bioengineering, vol. 11, no. 9, p. 899, sep. 2024, doi:
Novasinergia 2026, 9(2), 148-173 172
10.3390/bioengineering11090899.
[16] G. Bongiorno et al., “The Kinematic and Electromyographic Analysis of Roller Skating at Different
Speeds on a Treadmill : A Case Study,” Sensors, vol. 24, no. 17, p. 5738, sep. 2024, doi:
10.3390/s24175738.
[17] Z. Cao, G. Hidalgo, T. Simon, S. E. Wei, and Y. Sheikh, “OpenPose: Realtime Multi-Person 2D Pose
Estimation Using Part Affinity Fields,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 43, no. 1, pp. 172–
186, jan. 2021, doi: 10.1109/TPAMI.2019.2929257.
[18] C. Zheng et al., “Deep Learning-based Human Pose Estimation : A Survey,” ACM Comput. Surv., vol.
56, no. 1, pp. 1–37, aug. 2023, doi: 10.1145/3603618.
[19] W. Hsu et al., “PoseShot : hybrid CNN – BiLSTM transformer model for free throw action recognition
via pose analysis,” Sci Rep, vol. 16, no. 1, mar. 2026, doi: 10.1038/s41598-026-41025-0..
[20] S. C. Bakchy, M. Islam, M. R. Mahmud, and F. Imran, “Human Gait Analysis using Gait Energy
Image,” 2022, arXiv. doi: 10.48550/ARXIV.2203.09549.
[21] J. Loureiro and P. Lobato Correia, “Using a Skeleton Gait Energy Image for Pathological Gait
Classification,” in Proc. 15th IEEE Int. Conf. Autom. Face Gesture Recognition (FG 2020). IEEE, pp. 503–
507, nov. 2020. doi: 10.1109/fg47880.2020.00064.
[22] P. Lee, T. Chen, H. Lin, L. Yeh, C. Liu, and Y. Chen, “Integrating OpenPose and SVM for Quantitative
Postural Analysis in Young Adults : A Temporal-Spatial Approach,” Bioengineering, vol. 11, no. 6, p.
548, may 2024, doi: 10.3390/bioengineering11060548.
[23] Z. Liu, M. Yang, K. Li, and X. Qin, “Recognition of TaeKwonDo kicking techniques based on
accelerometer sensors,” Heliyon, vol. 10, no. 12, p. e32475, jun. 2024, doi: 10.1016/j.heliyon.2024.e32475.
[24] I-H. Chung, “Using break-even analysis to explore the cost and carbon reduction benefits of solar and
wind energy integration in microgrids for convenience stores,” Heliyon, vol. 10, no. 21, p. e39644, nov.
2024, doi: 10.1016/j.heliyon.2024.e39644.
[25] P. Jafarzadeh, P. Virjonen, P. Nevalainen, F. Farahnakian, and J. Heikkonen, “Pose Estimation of
Hurdles Athletes using OpenPose,” in 2021 International Conference on Electrical, Computer,
Communications and Mechatronics Engineering (ICECCME). IEEE, pp. 1–6, oct. 07, 2021. doi:
10.1109/iceccme52200.2021.9591066.
[26] M. E. Özateş, A. Yaman, F. Salami, S. Campos, S. I. Wolf, and U. Schneider, “Identification and
interpretation of gait analysis features and foot conditions by explainable AI,” Sci. Rep., vol. 14, no. 1,
mar. 2024, doi: 10.1038/s41598-024-56656-4..
[27] J. Stenum, C. Rossi, and R. T. Roemmich, “Two-dimensional video-based analysis of human gait using
pose estimation,” PLoS Comput Biol, vol. 17, no. 4, p. e1008935, Apr. 2021, doi:
10.1371/journal.pcbi.1008935.
[28] R. Hu et al., “Effective evaluation of HGcnMLP method for markerless 3D pose estimation of
musculoskeletal diseases patients based on smartphone monocular video,” Front. Bioeng. Biotechnol.,
vol. 11, jan. 2024, doi: 10.3389/fbioe.2023.1335251.
[29] N. J. Cronin et al., “Feasibility of OpenPose markerless motion analysis in a real athletics competition,”
Front. Sports Act. Living, vol. 5, jan. 2024, doi: 10.3389/fspor.2023.1298003.
[30] M. Mundt, “Bridging the lab-to-field gap using machine learning : a narrative review,” Sport. Biomech.,
vol. 24, no. 10, pp. 2779–2798, apr. 2023, doi: 10.1080/14763141.2023.2200749.
[31] A. Alzahrani and A. Ullah, “Advanced biomechanical analytics : Wearable technologies for precision
health monitoring in sports performance,” Digital Health, vol. 10, jan. 2024, doi:
10.1177/20552076241256745.
[32] D. R. Seshadri et al., “Wearable Technology and Analytics as a Complementary Toolkit to Optimize
Workload and to Reduce Injury Burden,” Front. Sports Act. Living, vol. 2, jan. 2021, doi:
10.3389/fspor.2020.630576.
[33] A. A. Phatak, F. G. Wieland, K. Vempala, F. Volkmar, and D. Memmert, “Artificial Intelligence Based
Body Sensor Network Framework — Narrative Review : Proposing an End ‑ to ‑ End Framework using
Wearable Sensors, Real ‑ Time Location Systems and Artificial Intelligence/Machine Learning
Algorithms for Data Collection , Data Mining and Knowledge Discovery in Sports and Healthcare,”
Sport. Med. - Open, vol. 7, no. 79, oct. 2021, doi: 10.1186/s40798-021-00372-0.
[34] M. Souaifi et al., “Artificial Intelligence in Sports Biomechanics : A Scoping Review on Wearable
Technology, Motion Analysis, and Injury Prevention,” Bioengineering, vol. 12, no. 8, p. 887, aug. 2025,
doi: 10.3390/bioengineering12080887.
Novasinergia 2026, 9(2), 148-173 173
[35] E. E. Cust, A. J. Sweeting, K. Ball, and S. Robertson, “Machine and deep learning for sport-specific
movement recognition: a systematic review of model development and performance,” vol. 37, no. 5,
pp. 568–600, oct. 2019, doi: 10.1080/02640414.2018.1521769.
[36] B. Van Hooren, N. Pecasse, K. Meijer, and J. M. N. Essers, “The accuracy of markerless motion capture
combined with computer vision techniques for measuring running kinematics,” Scand J Med Sci Sport,
vol. 33, no. 6, pp. 966–978, feb. 2023, doi: 10.1111/sms.14319.
[37] L. Xiang, A. Wang, Y. Gu, L. Zhao, V. Shim, and J. Fernandez “Recent Machine Learning Progress in
Lower Limb Running Biomechanics With Wearable Technology : A Systematic Review,” Front.
Neurorobot., vol. 16, jun. 2022, doi: 10.3389/fnbot.2022.913052.
[38] C. Dindorf et al., “From lab to field with machine learning – Bridging the gap for movement analysis
in real-world environments: A commentary,” Curr. Issues Sport Sci., vol. 9, no. 4, p. 014, sep. 2024, doi:
10.36950/2024.4CISS014.
[39] T. Y. Wang, J. Cui, and Y. Fan, “A wearable-based sports health monitoring system using CNN and
LSTM with self-attentions,” PLoS One, vol. 18, no. 10, p. e0292012, oct. 2023, doi:
10.1371/journal.pone.0292012.