1203
1 INTRODUCTION
The rapid development of machine learning
technology finds increasing application in waterborne
transport, particularly in vessel navigation. Potential
fields of applicability of AI methods can be seen in,
among other things, vessel traffic prediction, detection
or tracking [1-4]. Sensors acquiring image data are a
special case in performing the above tasks. Digital
images are ideal for automated acquisition of
information about navigation objects, as they can be
combined with deep learning methods to automate
navigation processes. Autonomous vehicles play an
important role in terms of target applicability [5,6],
shaping the direction of development of intelligent
solutions in shipping and often based on knowledge
obtained through image processing. Appropriate
implementation of artificial intelligence can
significantly improve the safety and efficiency of water
transport and prepare the ground for its
autonomisation. It should be noted in this connection
that the development of intelligent systems for IWT is
now firmly on the agenda, with plans for their
implementation extending to 2050 [7,8].
A special case of image processing applications
based on deep learning is inland navigation, where
navigation signs play a key role in the navigation
process. Their main role is to ensure safety and regulate
traffic rules on waterways. Misinterpretation of this
type of information can lead to serious incidents, such
as collisions with bridges - important pieces of
infrastructure [9]. Navigation signs are used in the
navigation process when operating a vessel and are
one of the elements of IENC (Inland Electronic
Navigational Chart). Hence, another potential
application of neural networks could be systems for
automatic mapping of navigation objects, where, in
addition to the detection efficiency itself, its accuracy
will be important. Automating this process could also
Automatic Detection of Navigational Signs on Inland
Waterways Using YOLO Neural Networks
P. Adamski & J. Łubczonek
Maritime University of Szczecin, Szczecin, Poland
ABSTRACT: The study analysed the detection of navigation signs for inland navigation using YOLO neural
networks. All major versions of the network available at the time of the study were analysed, i.e. from YOLO 1 to
YOLO 12. The study considered two criteria: detection efficiency and detection accuracy. The first case is related
to applications requiring the highest possible number of object detections, while the second is related to mapping
tasks, where the accuracy of determining the location of a sign plays an important role. The results of the study
showed that different efficiencies can be expected for the neural models studied. The latest models do not always
prove to be the best. In terms of detection efficiency, the YOLO 4 network proved to be the best model, while in
terms of sign detection precision, YOLO v7 had the highest horizontal accuracy and YOLO v10 had the highest
vertical accuracy. The results of the study indicated that in some cases, attention should be paid to the oldest
versions of YOLO, such as YOLO V1 and V2, which significantly reduce false detections.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 19
Number 4
December 2025
DOI: 10.12716/1001.19.04.18
1204
significantly speed up database updates [10] and
streamline the process of creating the maps themselves
[11], which often requires time-consuming indoor
work. As can be seen, in general navigation
applications, the most common use of neural networks
will be object detection. For navigation object mapping
systems, the best result achieved by a neural model will
be both high detection and accuracy. This will make it
possible to complete the largest number of objects as
well as to determine their exact position later on.
This paper focuses attention on the use of YOLO
convolutional networks, enabling real-time detection
[12]. This type of network is suitable both for
applications requiring ongoing sign detection (e.g. for
automated situation assessment based on detected
objects) and for mapping tasks. Due to the wide range
of applicability of this type of network, a comparative
analysis of all available major versions of the YOLO
network was carried out. The research focused on
identifying the most effective version, considering two
aspects: detection efficiency and detection accuracy.
The objects to be detected were standardised
navigation signs for inland navigation [13].
As can be seen, the problem of detecting fairway
markings has different requirements for the neural
network depending on the use case of the vision
system. A detection system for automated navigational
analyses, in which the priority action will be to detect
all the marks that are present in the fairway, requires
that the TP (True Positive) score is as high as possible,
even at the expense of a poorer FP (False Positive)
score. In such cases, the need to verify data is often less
of a problem for the user than omitting signs. However,
the system responsible for mapping signs has different
requirements. Such a system should be characterised
by the highest possible indication of IoU (Insertion
Over Union), which will allow the exact position of the
mark to be determined. In this case, the user knows that
the mark is already in the image, so the task of the
neural network is to indicate the position of the object
being searched for as accurately as possible. With such
a system, it is also desirable to achieve as low an FP as
possible, to reduce the need for manual filtering of data
after measurements have been taken.
2 METHODOLOGY
The research was conducted in two stages, related to
the assessment of the effectiveness of navigation
marking detection. In the first stage, the effect of YOLO
network type on object detection performance was
analysed. The following measures were used: TP (True
Positives), FP (False Positives), FN (False Negatives),
Precision, Recall, F1-score and average IoU
(Intersection over Union). The changes in Precision, TP
and FP parameters were then analysed as a function of
the IoU threshold. These values were examined in a
range from 0%, meaning that the detection has at least
one point in common with the label, to a value of 99%,
where this value means that the common area is almost
identical to the sum of the detection and label areas.
The aim of this step was to generally evaluate YOLO-
type neural networks in the context of detecting inland
waterway markings. In this case, the desired effect of
the neural network model is to maximise the number
of detected signs (TP) and the quality of detection, i.e.
the highest possible IoU.
The second stage of the research involved analysing
the precision of sign detection. The deviation of the
centre point of detection in the vertical and horizontal
planes from the position of the centre of the object
marked in the image in proportion to the length or
width of the object marked in the image, respectively,
was used as a measure of the precision of the detection
position. The calculation of the precision of the object's
position is shown in formulas 1 and 2.
100%, 100%
cc
wh
ll
xy
wh


= =
where:
w - horizontal precision of the detected object.
h - vertical precision of the detected object.
∆xc - horizontal distance of the centre point of the
detected object from the centre point of the label.
∆yc - vertical distance of the centre point of the detected
object from the centre point of the label.
wl – label width
hl - label height
2.1 Data
A total of 3269 images were prepared, with resolutions
ranging from 1920x1080 to 1920x1440. The
photographs will show inland navigation signs,
observed both from the quay and from the water. The
images show signs against a variety of backgrounds,
including bridges, buildings, coastal infrastructure
features and vegetation. Each photo contains between
1 and 15 manually labelled signs. The training
collection contains 2692 images, the validation
collection 194 images and the test collection 384
images. The photos show the photographed beacons
between 2008 and 2022 on the Odra River. Exemplary
photo is illustrated in fig. 1.
Figure 1. Example of a sign photo taken in 2009, West Oder
River, city of Szczecin
2.2 Models
The models were trained using the Darknet [14] and
Ultralytics [15] frameworks. The models: YOLOv1 [16],
YOLOv2 [17], YOLOv3 [18], YOLOv4 [19] and
YOLOv7 [20] were trained using configuration files
and software in the latest version of the Darknet
framework developed by Bochkovskii A. [21]. The
models were trained in the Ultralytics framework,
1205
YOLOv5 [22], YOLOv6 [23], YOLOv8 [24], YOLOv9
[25], YOLOv10 [26], YOLOv11 [27], YOLOv12 [28],
used the medium profile, labelled M, whose input size
is 640x640. All models were trained for 300 epochs,
using a default set of hyperparameters provided by the
framework and model configuration file. The models
used in the study, together with the input size, are
summarised in table 1.
Table 1. Models used in the study
Input size
416 x 416
416 x 416
416 x 416
608 x 608
640 x 640
640 x 640
640 x 640
640 x 640
640 x 640
640 x 640
640 x 640
640 x 640
2.3 Hardware and software
Model training was carried out on a desktop computer
with the configuration: Intel Core i9 14900K processor,
RAM: 128 GB DDR5 5400 MHz, graphics card: Nvidia
RTX 4090 24GB. Ubuntu 24.04 LTS was used as the
operating system and the drivers for the graphics card
- version 550.107.02 - were installed, along with the
CUDA framework [29] version 12.4. The Darknet
framework has been compiled from source code with
support for GPU computing in the CUDA framework.
The Ultralytics framework based on PyTorch [30] has
the ability to train using Nvidia GPUs as standard.
3 RESULTS
The first stage of the study analysed what the results of
standard neural network metrics such as the number of
TPs, FPs, FNs, Presicion, Recall, F1-score and mean IoU
were for the YOLO models at the settings: confidence
threshold: 0.5, NMS threshold: 0.4, IoU threshold 0.5.
The results are presented in Table 2.
Table 2. Results obtained for the different versions of YOLO.
Model
TP
FP
FN
Precision
Recall
F1-
score
Average
IoU
(darknet) YOLO v1
456
8
405
0,98
0,53
0,69
0,80
(darknet) YOLO v2
461
20
400
0,96
0,54
0,69
0,80
(darknet) YOLO v3
750
42
111
0,95
0,87
0,91
0,83
(darknet) YOLO v4
810
17
51
0,98
0,94
0,96
0,89
(darknet) YOLO v7
776
14
85
0,98
0,90
0,94
0,88
(ultralytics) YOLO v5
768
22
93
0,97
0,89
0,93
0,89
(ultralytics) YOLO v6
757
29
104
0,96
0,88
0,92
0,89
(ultralytics) YOLO v8
763
22
98
0,97
0,89
0,93
0,89
(ultralytics) YOLO v9
771
29
90
0,96
0,90
0,93
0,89
(ultralytics) YOLO v10
768
19
93
0,98
0,89
0,93
0,89
(ultralytics) YOLO v11
776
31
85
0,96
0,90
0,93
0,89
(ultralytics) YOLO v12
772
35
89
0,96
0,90
0,93
0,89
The highest TP score was achieved by the YOLO v4
model, which correctly detected 810 objects, achieving
a similar precision score to the YOLO v1, YOLO v7 and
YOLO v10 models. The differences in the precision of
the models are small, ranging from 0.95 to 0.98.
Significantly greater differences can be observed in the
Recall and F1-score indicators. Recall for all networks
ranged from 0.53 to 0.94. Clearly worse Recall scores
were obtained by the oldest models YOLOv1 - 0.53 and
YOLOv2 - 0.54, while the scores of all the others were
similar and ranged from 0.87 to 0.94. The F1-score
results are similar, with the YOLOv1 and YOLOv2
models performing noticeably worse (at 0.69), while
the results of the other models ranged from 0.91 to 0.96.
The highest F1-score was achieved by the YOLOv4
model. The average IoU for all networks ranged from
0.8 to 0.89, with only the score of models YOLOv1 to
YOLOv3 slightly underperforming compared to the
others. The lowest FP score was achieved by model
YOLOv1, which detected 8 unwanted objects.
It was also analysed how precision varies with the
IoU Threshold set, with confidence threshold: 0.5 and
NMS threshold: 0.4, the results are presented in the
graph in Figure 2.
Figure 2. Graph of the dependence of the Precision indicator
on the set IoU Threshold
For IoU threshold values between 0 and 0.5, all
models achieved similar precision. Only above an IoU
threshold of 0.5 do the models YOLOv1, YOLOv2 and
YOLOv3 begin to show significantly lower precision.
Figure 3 shows the dependence of TP on IoU
threshold. The model with the best characteristics is
YOLOv4, which achieves a significantly higher TP than
the other models in the range from 0.0 to 0.85 IoU
Threshold. The YOLO v3 model in the IoU Threshold
range from 0 to 0.45 achieves a similar result to the
other models, but above this range it shows a
significantly lower TP.
Figure 3. Diagram of the dependence of the TP indicator on
the IoU threshold.
The dependence of FP on IoU threshold was also
tested (Figure 4). The YOLO v1 and YOLO v2 models
proved to be interesting cases. For IoU threshold
values between 0.0 and 0.5, where they showed the
1206
lowest amount of FP, while for IoU threshold values
between 0.55 and 0.9 they showed a significant increase
in FP, clearly greater than most of the other models.
Figure 4. Graph of the dependence of the FP indicator on the
IoU threshold.
It follows that these models detect the position of
objects in the image with significantly lower accuracy
than newer models such as YOLOv7, YOLOv4, and all
models in the Ultralytics framework. Only the YOLO
v3 model performs worse than YOLO v1 and YOLO v2
for a high (above 0,5) IoU threshold. An example of
detection accuracy is shown in Figure 5.
Figure 5. Detection results for the YOLO v1 (left) and
YOLOv7 (right) networks. Red shows the prediction made by
the model, blue the ground truth.
The second stage of the research involved analysing
the precision of the prediction centre point. The
measurement of the precision of the prediction centre
point indicates that the first three models – YOLOv1,
YOLOv2 and YOLOv3 – are characterised by
significantly lower prediction precision compared to
newer versions. In the case of these models, lower
precision in the horizontal axis than in the vertical axis
is also noticeable. The remaining models achieved very
similar results. These data are presented in Figure 6.
Figure 6. Average relative precision of prediction centre
determination
Table 3 presents the values of the average and
maximum relative prediction offset in relation to its
size for all analysed models. The largest maximum
shift in the horizontal axis was observed for the YOLO
v6 - 108.4% and YOLO v11 – 105.9 % models. These
models obtained a maximum error significantly higher
than the others. The lowest maximum horizontal
position error was achieved by the YOLO v7 model,
which was 22.2%. The YOLO v1, YOLO v2 and
YOLOv3 models achieved lower horizontal position
precision than the other models, at 6%, 6.4% and 4.2%
respectively. The other models achieved similar
precision, ranging from 2% to 2.5%. The average
vertical position measurement results for the YOLO v1,
YOLO v2, and YOLO v3 models were less different
from the other models, at 3.4%, 3.2%, and 2.5%,
respectively, while all results ranged from 1.9% to
3.4%.
Table 3. Values of the mean and maximum relative offset of
the prediction relative to its size for all models analysed
Model
Average
Horizontal
Offset
Maximum
Horizontal
Offset
Average
Vertical
Offset
Maximum
Vertical
Offset
YOLOv1
6,0%
60,7%
3,4%
61,3%
YOLOv2
6,4%
75,2%
3,2%
24,0%
YOLOv3
4,2%
52,6%
2,5%
70,3%
YOLOv4
2,3%
50,0%
1,9%
27,5%
YOLOv5
2,1%
30,5%
2,3%
32,4%
YOLOv6
2,5%
108,4%
2,5%
81,2%
YOLOv7
2,2%
22,2%
2,1%
68,5%
YOLOv8
2,1%
54,8%
2,3%
19,1%
YOLOv9
2,2%
83,3%
2,4%
38,2%
YOLOv10
2,2%
64,5%
2,3%
15,9%
YOLOv11
2,3%
105,9%
2,5%
120,7%
YOLOv12
2,0%
36,1%
2,3%
19,6%
In addition, the results of maximum shifts are
presented in Figure 7. The lowest maximum errors in
determining the vertical position were obtained by the
YOLO v8 and YOLO v12 models, with offset values of
19.1% and 19.6%, respectively. The worst result was
achieved by the YOLOv12 model – 120.7%.
Figure 7. results of maximum shifts of the prediction centre
point
A maximum prediction centre point shift of more
than 100% means that the prediction centre is further
away than the dimension of the label applied to the
image. Such a case is presented in Figure 8, where a
horizontal and vertical shift of the centre point of
prediction by a value greater than the width and height
of the label applied to the image was observed.
1207
Figure 8. YOLO v11 prediction with vertical and horizontal
shift of more than 100%
4 DISCUSSION
The choice of the best model may depend on the
specific task. In the case where the main problem is to
detect all objects, e.g. during automated navigation
analyses, the best YOLO model with this training and
test data set turned out to be YOLOv4, which had the
highest TP score - a model that is already quite old,
given how many newer models are currently available.
For a task where the most important element is to
reduce the number of erroneous detections required
for subsequent manual filtering, YOLOv1, the first
YOLO network model to achieve an FP of 8, proved to
be the best. However, if the most important task
performed by the network is to accurately indicate the
position of the searched object in the image, then
YOLOv10 proved to be the best model, achieving the
highest precision at an IoU threshold of 0.95. This
means that this model most accurately indicates the
position of the searched objects. The difference
between the models from YOLO v4 to YOLO v12 is not
significant, and only the oldest models, YOLO v1,
YOLO v2 and YOLO v3, deviate from these results.
Newer models are not always better in terms of the
quality of the data they provide – this is particularly
evident in the results of the YOLOv4 and YOLOv7
models trained in the Darknet framework. These
models can still be useful in many applications, and
upgrading to a newer model is not always justified.
The oldest models, YOLOv1 and YOLOv2, lag
significantly behind the others, especially in terms of
the number of TP objects detected, but they can still be
useful in certain specific cases. The latest available
model, YOLOv12, achieved good results, although in
many tasks it proved to be inferior to older models.
Determining the centre point of an object in an
image may require the use of a model that is closely
tailored to the specific task. For example, in systems
whose sole purpose is to determine the horizontal
position of objects, YOLOv5 or YOLOv7 models may
be the appropriate choice. In the measurements carried
out, they not only achieved a low average error in
determining the centre point of prediction, but also the
lowest values of maximum shifts in the horizontal axis.
In contrast, for applications requiring precise
vertical positioning - such as measuring the height of
signage location relative to the observer - the YOLOv8
model would be the best choice. It showed the lowest
maximum error in the vertical position of the
prediction centre point, while also having a low
average error.
5 CONCLUSIONS
Older versions of the YOLO models, such as YOLOv4
and YOLOv7, despite the advent of newer
architectures, still prove to be useful in the context of
detection of fairway markings. The results obtained by
these models are at a level comparable to the newer
models.
The oldest models – YOLOv1 and YOLOv2 – differ
in terms of quality from newer models. Their
architectures are outdated compared to current
standards, which means that they do not perform well
in detecting objects in most cases. Their low precision
and limited ability to locate signs make them
unsuitable for most applications in navigation sign
detection. In certain specific cases, the oldest models
– YOLO v1 and YOLO v2 – can be useful when the
essence of the problem is to limit FP.
The highest precision in determining horizontal
position was achieved by the YOLO v7 model, while
the highest precision in determining vertical position
was achieved by the YOLO v10 model.
Models from YOLO v4 to YOLO v12 achieved
similar results of average precision in determining the
centre point of an object, the choice of the appropriate
model in this case can be made based on other factors.
In terms of the number of objects detected in the
image, the YOLOv4 model achieved the best results.
This model is best suited for tasks where it is necessary
to detect as many navigation signs as possible.
REFERENCES
[1] Azimi, S.; Salokannel, J.; Lafond, S.; Lilius, J.; Salokorpi,
M.; Porres, I. A Survey of Machine Learning Approaches
for Surface Maritime Navigation. 2020. Available online:
http://hdl.handle.net/2117/329714 (accessed on 12
February 2025).
[2] Donandt, K.; Böttger, K.; Söffker, D. Short-Term Inland
Vessel Trajectory Prediction with Encoder-Decoder
Models. In Proceedings of the 2022 IEEE 25th
International Conference on Intelligent Transportation
Systems (ITSC), Macau, China, 8–12 October 2022; pp.
974–979. https://doi.org/10.1109/ITSC55140.2022.9922148.
[3] Agorku, G.; Hernandez, S.; Falquez, M.; Poddar, S.;
Amankwah-Nkyi, K. Traffic Cameras to Detect Inland
Waterway Barge Traffic: An Application of Machine
Learning, Computer Vision and Pattern Recognition.
arXiv 2024, arXiv:2401.03070.
[4] Hart, F.; Okhrin, O.; Treiber, M. Vessel-Following Model
for Inland Waterways Based on Deep Reinforcement
Learning. Ocean. Eng. 2023, 281, 114679.
https://doi.org/10.1016/j.oceaneng.2023.114679.
[5] Vanneste, A.; Vanneste, S.; Vasseur, O.; Janssens, R.;
Billast, M.; Anwar, A.; Mets, K.; De Schepper, T.; Mercelis,
S.; Hellinckx, P. Safety Aware Autonomous Path
Planning Using Model Predictive Reinforcement
Learning for Inland Waterways. In Proceedings of the
IECON 2022–48th Annual Conference of the IEEE
Industrial Electronics Society, Brussels, Belgium, 17–20
October 2022; pp. 1–6.
https://doi.org/10.1109/IECON49645.2022.9968678.
1208
[6] Qiao, Y.; Yin, J.; Wang, W.; Duarte, F.; Yang, J.; Ratti, C.
Survey of Deep Learning for Autonomous Surface
Vehicles in Marine Environments. IEEE Trans. Intell.
Transp. Syst. 2023, 24, 3678–3701.
https://doi.org/10.1109/TITS.2023.3235911.
[7] Hao, G.; Xiao, W.; Huang, L.; Chen, J.; Zhang, K.; Chen, Y.
The Analysis of Intelligent Functions Required for Inland
Ships. J. Mar. Sci. Eng. 2024, 12, 836.
https://doi.org/10.3390/jmse12050836.
[8] Li, Y.; Hu, Y.; Rigo, P.; Lefler, F.E.; Zhao, G.; Eds.
Proceedings of PIANC Smart Rivers 2022: Green
Waterways and Sustainable Navigations; Lecture Notes
in Civil Engineering; Springer: Singapore, 2023; Volume
264. https://doi.org/10.1007/978-981-19-6138-0.
[9] Fan, W.; Zhong, Z.; Wang, J.; Xia, Y.; Wu, H.; Wu, Q.; Liu,
B. Vessel-Bridge Collisions: Accidents, Analysis, and
Protection. China Journal of Highway and Transport.
2024, 37(5), 38–66.
[10] Łubczonek, J., i M. Włodarczyk. Wykorzystanie geobazy
danych w procesie tworzenia elektronicznych map
nawigacyjnych dla żeglugi śródlądowe [Application of
geodatabase in the process of creation electronic
navigational charts for inland shipping]. Archiwum
Fotogrametrii, Kartografii i Teledetekcji, t. 21, 2010, s.
221–34.
[11] Łubczonek, J. Opracowanie i implementacja
elektronicznych map nawigacyjnych dla systemu RIS w
Polsce [Elaboration and implementation of electronic
navigational charts for RIS in Poland]. Roczniki
Geomatyki 2015, 13, 359–368.
[12] Adamski, P.; Lubczonek, J. A Comparative Analysis of
the Usability of Consumer Graphics Cards for Deep
Learning in the Aspects of Inland Navigational Signs
Detection for Vision Systems. Appl. Sci. 2025, 15, 5142.
https://doi.org/10.3390/app15095142
[13] SIGNI. European Code for Signs and Signals on Inland
Waterways: Resolution No. 90; United Nations: New
York, NY, USA, 2018.
[14] Redmon J. Darknet: Open Source Neural Networks in C.
Online source: https://pjreddie.com/darknet. Access date
13.07.2025
[15] Jocher G.; Qiu J.; Charasia A. Ultralyiics YOLO. Online
source: https://github.com/ultralytics/ultralytics Access
date:13.07.2025
[16] Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. You
only look 1182 once: Unified, real-time object detection.
In Proc. IEEE Conf. Comput (Vol. 1183, pp. 779-788).
[17] Redmon, J., Farhadi, A. YOLO9000: Better, Faster,
Stronger. Proceedings of the IEEE conference on
computer vision and pattern recognition
[18] Farhadi A., Redmon J. "Yolov3: An incremental
improvement." In Computer vision and pattern
recognition, vol. 1804, pp. 1-6. Berlin/Heidelberg,
Germany: Springer, 2018.
[19] Bochkovskiy, Alexey & Wang, Chien-Yao & Liao, Hong-
yuan. (2020). YOLOv4: Optimal Speed and Accuracy of
Object Detection. 10.48550/arXiv.2004.10934.
[20] Wang, C., Bochkovskiy, A., & Liao, H. M. (2022).
YOLOv7: Trainable bag-of-freebies sets new state-of-the-
art for real-time object detectors. arXiv (Cornell
University). https://doi.org/10.48550/arxiv.2207.02696
[21] Bochkovskiy, A. Open Source Neural Networks in C.
Online source: https://github.com/AlexeyAB/darknet
Access date:13.07.2025
[22] Khanam, R., & Hussain, M. (2024). What is YOLOv5: A
deep look into the internal features of the popular object
detector. arXiv (Cornell University).
https://doi.org/10.48550/arxiv.2407.20892
[23] Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z.,
Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y.,
Zhou, L., Xu, X., Chu, X., Wei, X., & Wei, X. (2022).
YOLOV6: A Single-Stage Object Detection Framework for
Industrial Applications. arXiv (Cornell University).
https://doi.org/10.48550/arxiv.2209.02976
[24] R. Varghese and S. M., "YOLOv8: A Novel Object
Detection Algorithm with Enhanced Performance and
Robustness," 2024 International Conference on Advances
in Data Engineering and Intelligent Computing Systems
(ADICS), Chennai, India, 2024, pp. 1-6, doi:
10.1109/ADICS58448.2024.10533619.
[25] Wang, C., Yeh, I., & Liao, H. M. (2024). YOLOV9:
Learning what you want to learn using programmable
gradient information. arXiv (Cornell University).
https://doi.org/10.48550/arxiv.2402.13616
[26] Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., &
Ding, G. (2024). YOLOV10: Real-Time End-to-End Object
Detection. arXiv (Cornell University).
https://doi.org/10.48550/arxiv.2405.14458
[27] Khanam, R., & Hussain, M. (2024b). YOLOV11: An
overview of the key architectural enhancements. arXiv
(Cornell University).
https://doi.org/10.48550/arxiv.2410.17725
[28] Tian, Y., Ye, Q., & Doermann, D. (2025). YOLOV12:
Attention-Centric Real-Time Object Detectors. arXiv
(Cornell University).
https://doi.org/10.48550/arxiv.2502.12524
[29] Nvidia Corporation. CUDA Toolkit Documentation 12.4
Available online:
https://docs.nvidia.com/cuda/archive/12.4.0/ (accessed
on 13 July 2025).
[30] Paszke, A.; et al. PyTorch: An Imperative Style, High-
Performance Deep Learning Library. arXiv 2019,
arXiv:1912.01703. Available online:
http://arxiv.org/abs/1912.01703 (accessed on 12 February
2025)