IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 2, April 2025, pp. 945∼953

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i2.pp945-953 ❒ 945

Security in smart cities using YOLOv8 to detect lethal

weapons
Ederson Rodriguez-Rosas, Aron Castillo-Turpo, Kevin Acuña-Condori, Ernesto Paiva-Peredo
Department of Electronic Engineering, Faculty of Engineering, Universidad Tecnológica del Perú, Lima, Peru

Article Info ABSTRACT

Article history: The increase in the illegal use of lethal weapons at a global level has reached
worrying figures, resulting in an increase in assaults and armed robberies. Based
Received Mar 16, 2024
on the above, closed circuit television (CCTV) surveillance systems emerge as
Revised Nov 14, 2024 an alternative solution. Therefore, the use of artificial intelligence is explored in
Accepted Nov 24, 2024 order to detect the presence of lethal weapons in images accurately. In this study,
a convolutional neural network model YOLOv8 is trained. A database including
Keywords: 4104 images with the presence of lethal weapons is generated. The Google
Colab platform is used for the training phase, since it provides us with a free
Convolutional neural networks graphic processing unit (GPU), and the YOLOv8x and YOLOv8n models are
Detection bladed weapons used for comparison. The results indicate an advantage when using the YOLOv8
Google Colab models, and when comparing them with similar models already proposed in the
Weapons detection studied literature, we can conclude that our model stands out with an accuracy
YOLOv8 of 89.56% in the detection of lethal weapons. In conclusion, we were able to
obtain a model capable of detecting lethal weapons in CCTV images, in addition
to being able to be used in applications that require real-time detection.

This is an open access article under the CC BY-SA license.

Corresponding Author:
Ernesto Paiva-Peredo
Department of Electronic Engineering, Faculty of Engineering, Universidad Tecnológica del Perú
125 Natalio Sanchez Street, Santa Beatriz Urbanization, Cercado de Lima, Lima, Peru
Email: [email protected]

1. INTRODUCTION
Globally, there are more than 1,013 million firearms in the world and more than 85% of them are in
the hands of civilians [1]. Consequently, the use of firearms causes up to 1,000 deaths per day, along with
more than 250,000 armed incidents per year [2]. This reality highlights the importance of the use of closed
circuit television (CCTV) systems for the early identification of lethal weapons (firearms and sharp weapons)
in images [3]–[9]. In this way, it is possible to combat crimes such as assaults and robberies, which are carried
out with armed hands [5], [10]. In this context, the detection of crimes by images based on artificial intelligence
arises, due to its high precision, anticipation and adaptability in the detection of objects [6], [11].
Currently, the use of artificial intelligence, for the identification of violence scenarios such as crimes
involving the presence of lethal weapons, is in constant expansion. Therefore, algorithms based on an artificial
neural network and the moving picture experts group 7 (MPEG-7) descriptor are proposed to classify frames
of a CCTV transmission, in which results of 94% and 95%, respectively, are obtained [4], [12]. It is relevant to
note that both models exhibit a low rate of positive incorrect detections, but miss a considerable amount of neg-
ative incorrect detections. However, among the most outstanding approaches is the application of convolutional
neural networks (CNN) [9], [10], [13]. Therefore, a network based on visual geometry group-16 (VGG-16)
and another on VGG were configured, which have an efficiency of 93% and 86%, respectively [10], [13]. It

Journal homepage: http://ijai.iaescore.com

946 ❒ ISSN: 2252-8938

is important to note that both models use the VGG extractor, which does not allow optimal performance when
locating the presence of lethal weapons in the images. In this sense, for the identification of a crime scene it is
important to detect lethal weapons, for this purpose, CNNs are usually used [3], [14]. For example, the method
called flow gated network, which combines the advantages of three-dimensional convolutional neural networks
(3D-CNN) and optical flow, resulting in an accuracy of 87.25% [3]. It is worth mentioning that 3D-CNNs
require a higher computational load with respect to CNNs [3].
First, in the field of CNNs, YOLO is one of the most widely used to identify objects in the frames of a
real-time video sequence [8]. For this reason, a YOLOv3 model was trained, with the purpose of identifying the
presence of firearms in images [2], [15], [16]. This model is fused with recurrent convolutional neural networks
(R-CNN), through which it achieves a performance of 94.23% [15]. It should be noted that the model uses the
open source database ”kaggle”, this generates a limited management in the quality and variety of the training
data to the network. It is also important to consider that the R-CNN does not work in real time [15]. Also,
models for weapon detection are proposed by applying YOLOv5 [1], [8], [17], [18]. The model exhibits the
ability to identify lethal weapons, achieving a mAP of 52.92%. It also exhibits an inference rate of 61 frames
per second (FPS) [8]. In addition, the database applied to the training of the network, is deficient based on the
variety. Also, an efficiency of 93% image accuracy was obtained by combining YOLOv5 and faster R-CNN,
using a database of 3000 guns [1].
On the other hand, for the recognition of firearms in images by means of deep learning, the fusion
between posture estimation and object detection is being used [5], [6]. Therefore, an algorithm is designed to
define the pose of each person in a frame, in order to obtain the position of the hands and create a bounding
box where the object detector is applied [5], [6]. The model uses Open Pose to estimate the posture and vision
transformer (ViT) to detect the weapon [5]. It is important to point out that the efficiency relies on the pose
estimator, if it fails, there will be no detection process [5], [6]. To achieve an early detection of firearms, it is
essential to consider that these are not always carried in the hands, since there are varieties of weapons that can
be carried on the chest or hanging from the shoulders.
Given the need for more efficient systems for the detection of dangerous objects, whether firearms or
knives. A YOLOv8 CNN is trained using a cloud computing infrastructure to reduce the computational burden
[19]. In addition, a diverse dataset is generated that includes a wide range of lethal weapons, such as shotguns,
pistols, knives, machetes, among others. To augment the training data, synthesis techniques are employed, this
increases the database and gives better learning versatility to the CNN [3], [5].

2. METHODOLOGY
Today, the number of crimes involving lethal weapons (firearms and knives) has risen exponentially
[1], [4]. Measures have been taken to address this problem, such as the installation of CCTV systems. How-
ever, these systems only accumulate the data, and do not work it through video inspection or object detection
algorithms [7], [9], [12]. Therefore, a neural network capable of detecting lethal weapons in real time is trained,
in order to provide greater accuracy, versatility and reduction of false positives. After conducting a comparative
analysis of various object detection techniques that rely on deep learning, such as YOLO, single-shot multibox
detector (SSD) and the fastest R-CNN. It is determined that YOLO excels in achieving an optimal trade-off
between mean average precision (mAP) and inference speed, for real-time predictions [8], [11], [18].
Finally, YOLOv8 is defined as the CNN to be trained, because it is the most recent iteration in a
sequence of algorithms created by Open AI researchers for object detection and tracking. To our knowledge,
this is the first work investigating the use of YOLOv8 in identifying deadly weapons in images. The research is
divided into three stages, the first one covers the obtaining of the database, the second stage corresponds to the
data augmentation together with the etiquedata and the third stage the training of the CNN YOLOv8, obtained
from ultralytics.

2.1. Data acquisition

In this stage, the collection of 3104 images is performed, proportionally 1554 images with the presence
of firearms and 1550 images with the presence of white weapon. Considering the importance of the variety of
data for a good training, we considered images of the different types of firearms and knives that exist [9], [10].
A collage of the collected images is presented in Figure 1. It is relevant to note that the dataset was obtained
from three sources, in order to have variety in the origin of the images. From Kaggle dataset 1040 images

Int J Artif Intell, Vol. 14, No. 2, April 2025: 945–953

Int J Artif Intell ISSN: 2252-8938 ❒ 947

were extracted, from Youtube 931 images were collected and from Google 1133 images were collected. The
acquisition process in the sources used is explained as follows.

Figure 1. Collage of collected images

2.1.1. Google images

A Google extension called ”download all images” was used. This tool allows us to download all the
images that have been visually loaded, and gives us a file in rar format, where all the images are in jpg format
[2]. In this way, we searched for images using keywords such as: people with knives, assaults with knives,
people with axes, people with firearms, assault with firearms.

2.1.2. YouTube
Additionally, the YouTube platform was used for the collection of images with the presence of a
sharp weapon, in this platform we can find videos of CCTV [3]. A flowchart is presented in Figure 2, of the
procedure of obtaining images from a video. Initially, we explored the YouTube platform using a series of
keywords related to violent acts, such as real knife fights, knife-wielding assailant, and other similar terms.
Subsequently, we used the online video converter program that automatically downloads the videos from the
obtained links. After completing this process, we extract images every 5 seconds, all at a speed of 30 FPS.

Figure 2. Video frame acquisition process

2.1.3. Kaggle dataset Online

The Kaggle platform hosts a wide range of datasets. These datasets can be used for machine learning
[17], [20], [21]. For the detection of lethal weapons, this online database was used to extract images whose
content contains any type of firearms or sharp weapons [15]. The justification for using Kaggle is the presence
of images obtained from a CCTV system, in addition to the variety of images presented by [15], [22].

2.2. Increased data and labeling

In order to increase the number of images collected, which were 3104, and also to improve the
variety of the database in terms of tonalities and visualization perspective, we propose the application of
data augmentation techniques [3], [5]. These techniques allow us to identify patterns in our data by performing
various transformations on the images, such as flipping, rotating, scaling, converting to black and white, among
others [14]. In this study we chose to change the image to gray tone and invert it horizontally. Figure 3 shows
the algorithm performed in the MATLAB environment, where a group of 500 images selected from the group
of images initially collected were loaded and the rgb2gray and fliplr commands were applied. Through this
technique we were able to increase 1000 more images and obtain a dataset of 4104 images with the presence
of lethal weapons.

Figure 3. Data augmentation process

Security in smart cities using YOLOv8 to detect lethal weapons (Ederson Rodriguez-Rosas)
948 ❒ ISSN: 2252-8938

On the other hand, the images are tagged in the Makesense online platform. For this purpose, two tags
”Firearm” and ”White weapon” were created. These labels provide the neural network with information about
the exact location of the object to be identified, so that the network learns to recognize patterns and relevant
features in the area limited by the label. Finally, the images are divided into two groups, 4104 for training and
892 for validation. Figure 4 shows a group of labeled images.

Figure 4. Collage of labeled images

2.3. Deep learning model YOLOv8

YOLO algorithms in computer vision are recognized for their high accuracy and small model sizes,
making them easy to use for both developers and machine learning experts [18]. The latest version, YOLOv8,
is used in a variety of applications, including object detection, image classification and segmentation [23].
Ultralytics has developed YOLOv8 as an evolution of YOLOv5, incorporating architectural improvements. A
distinguishing feature of YOLOv8 is that it does not use anchors; instead, it explicitly estimates the center
of objects, leading to a reduction in the number of predictions and speeding up the non-maximum suppression
(NMS) process [18]. YOLOv8 is broken down into five models, YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l,
and YOLOv8x, each with variations in accuracy and speed [11], [18], [23].
For this study, the YOLOv8x model is used, due to the results obtained in each parameter presented
in the table. Also, it is important to make a comparison between the YOLOv8x and YOLOv8n model, because
a wide difference is observed in the results using a COCO dataset. But we do not know their behaviors when
trained with our lethal weapon database [23].

2.4. YOLOv8 training

For training and validation of YOLOv8, we use Google Colab, a free cloud runtime environment, this
platform, provides us with a free online GPU [17]. Using a GPU significantly speeds up both the rendering
process and the compile time. When working with a large data set that is not limited only to numerical values,
there is an improvement in performance when running numerous steps [17].
The previously created database is uploaded to the Google Colab environment in ”rar” format, in
addition to a dot yaml file, where the paths of each of the two folders of the training and validation database are
found, also in this file the number of items to detect in this case firearms and sharp weapons is defined. Then,
the training is started by importing the ultralytics libraries and defining the YOLO version and model to be
used. Finally, the training parameters are defined, such as the number of epochs to be trained, the batch size to
be used and the size of the images to be processed. It should be noted that for the study a comparison between
two versions of YOLOv8 will be made, using YOLOv8n and YOLOv8x, due to the high margin of results in
each of its metrics obtained in the test performed by ultralytics.

2.5. Evaluation metrics

These are very important parameters to measure the performance of the model in terms of accuracy
and speed. This paper uses accuracy, detection rate, false detection rate and mAP as metrics to evaluate the
performance of the algorithm. Accuracy measures the ratio of correct detections to the total number of correct
detections [24]–[26]. While the recall parameter measures the ratio between the positive objects that have been
detected and the total number of positive objects [23]–[25], [27]. Finally, the mAP is widely used because it
provides information about the accuracy of all classes of objects present in the database [25], [27], [28].

3. RESULTS AND DISCUSSION

The training results of both YOLOv8 models are very close when observing their evaluation metrics.
Regarding the performance of the proposed YOLOv8x model for the detection of lethal weapons we obtained a
mAP of 87.8% and a mAP50-95 of 63.47%. In the identification of firearms we obtained a mAP of 88.7%, and
in the identification of sharp weapons a mAP of 86.8%, training the network for 50 epochs and with a batch
size of 16. The results of the evaluation metrics are shown in Figure 5.

Int J Artif Intell, Vol. 14, No. 2, April 2025: 945–953

Int J Artif Intell ISSN: 2252-8938 ❒ 949

Figure 5. YOLOv8x model results

On the other hand, the results of the YOLOv8n model with the same training parameters were very
similar to the previous one. Because a mAP50 of 89.59% and a mAP50-95 of 63.26% were obtained in the
detection of lethal weapons. This allows us to conclude that the YOLOv8n model has a better performance in
the detection of lethal weapons, this is observed when analyzing each evaluation metric in Figure 6.

Figure 6. YOLOv8n model results

Our YOLOv8n network demonstrates superior performance compared to existing methods in the
literature. The flow gated network, which combines 3D-CNN and optical flow for violence scene detection,
achieved a mAP of 87.25% [3]. In comparison, our YOLOv8n model achieved a higher mAP of 89.59%,
with lower computational requirements. Additionally, while a YOLOv3 and R-CNN hybrid model for firearm
detection reported a mAP of 85% [15], it is not suitable for real-time use due to its slower speed. Our YOLOv8n
model, with a mAP of 89.59%, supports real-time operation, making it ideal for early weapon detection.
Furthermore, it outperforms a YOLOv5 model designed to detect lethal weapons, which achieved a mAP
of 52.92% [8], thanks to our more diverse training dataset.

Security in smart cities using YOLOv8 to detect lethal weapons (Ederson Rodriguez-Rosas)
950 ❒ ISSN: 2252-8938

3.1. Confusion matrix

The confusion matrix is a critical tool for evaluating the performance of classification models.
Figure 7 presents the confusion matrices for the YOLOv8n and YOLOv8x models, enabling a comparative
analysis of their performance. In Figure 7(a), the confusion matrix for the YOLOv8n model is displayed,
illustrating the classification errors and the distribution of predicted versus actual labels. Figure 7(b) shows the
confusion matrix for the YOLOv8x model, providing a similar analysis and facilitating a comparison with the
YOLOv8n model.

(a)

(b)
Figure 7. Matrix of confusion for (a) the YOLOv8n model and (b) the YOLOv8x model

3.2. Prediction
The prediction results of YOLOv8n and YOLOv8x models are depicted in Figures 8 and 9. Figure 8
includes the prediction results for a batch of 16 images using both models. In Figure 8(a), the predictions of
the YOLOv8n model are shown, detailing how the model performs on the given images. Figure 8(b) presents
the prediction results of the YOLOv8x model for the same batch of images, allowing for a direct comparison
with the YOLOv8n model.
Figure 9 further illustrates the prediction performance of both models. Figure 9(a) highlights an
instance where the YOLOv8n model misidentifies a white weapon as a person’s cap, providing insight into
the model’s limitations. In contrast, Figure 9(b) shows the YOLOv8x model’s ability to correctly identify the

Int J Artif Intell, Vol. 14, No. 2, April 2025: 945–953

Int J Artif Intell ISSN: 2252-8938 ❒ 951

white weapon with high accuracy, even when only a partial view of the weapon is visible. This comparison
underscores the YOLOv8x model’s superior detection capabilities in specific scenarios.

(a) (b)
Figure 8. Prediction results of the (a) YOLOv8n model and (b) YOLOv8x model

(a) (b)
Figure 9. Detection performance comparison of the (a) YOLOv8n mislabels a weapon and (b) YOLOv8x
labels it correctly

4. CONCLUSION
In this study, we developed and evaluated a YOLOv8 network specialized in the detection of lethal
weapons, covering both firearms and edged weapons. The results obtained were exceptional, with a perfor-
mance rate of 89.56%. This figure validates the effectiveness of our approach and highlights the model’s ability
to accurately identify potential threats. In addition, we note that our YOLOv8n model exhibits superiority in
certain aspects compared to YOLOv8x, especially in terms of accuracy. However, it is essential to emphasize
that the prediction results do not allow us to state with certainty a general superiority, indicating the need for
more detailed analyses in future research. Our research work provides a trained network with the ability to be
implemented in real time in any CCTV system that has the necessary computational parameters, as would be
the case in a smart city. Also, we believe that our proposed model could enhance its performance by training
with a more diverse database. It is worth noting the importance of the size of the images used in the training,
since YOLOv8 operates with a recommended size of 640 pixels for image.

REFERENCES
[1] A. H. Ashraf et al., “Weapons detection for security and video surveillance using cnn and YOLO-v5s,” Computers, Materials and
Continua, vol. 70, no. 2, pp. 2761–2775, 2022, doi: 10.32604/cmc.2022.018785.
[2] S. Narejo, B. Pandey, D. E. Vargas, C. Rodriguez, and M. R. Anjum, “Weapon detection using YOLOv3 for smart surveillance
system,” Mathematical Problems in Engineering, vol. 2021, 2021, doi: 10.1155/2021/9975700.
[3] M. Cheng, K. Cai, and M. Li, “RWF-2000: an open large scale video database for violence detection,” in 2020 25th International
Conference on Pattern Recognition (ICPR), IEEE, 2021, pp. 4183–4190, doi: 10.1109/ICPR48806.2021.9412502.
[4] M. Grega, S. Lach, and R. Sieradzki, “Automated recognition of firearms in surveillance video,” in 2013 IEEE International Multi-

Security in smart cities using YOLOv8 to detect lethal weapons (Ederson Rodriguez-Rosas)
952 ❒ ISSN: 2252-8938

Disciplinary Conference on Cognitive Methods in Situation Awareness and Decision Support (CogSIMA), IEEE, 2013, pp. 45–50,
doi: 10.1109/CogSIMA.2013.6523822.
[5] J. Ruiz-Santaquiteria, A. Velasco-Mata, N. Vallez, O. Deniz, and G. Bueno, “Improving handgun detection through a combination
of visual features and body pose-based data,” Pattern Recognition, vol. 136, 2023, doi: 10.1016/j.patcog.2022.109252.
[6] J. Ruiz-Santaquiteria, A. Velasco-Mata, N. Vallez, G. Bueno, J. A. Alvarez-Garcia, and O. Deniz, “Handgun detection using com-
bined human pose and weapon appearance,” IEEE Access, vol. 9, pp. 123815–123826, 2021, doi: 10.1109/ACCESS.2021.3110335.
[7] E. Paiva-Peredo, A. Vaghi, G. Montù, and R. Bucher, “Human detection on antistatic floors,” SSRN Electronic Journal, 2022, doi:
10.2139/ssrn.4264059.
[8] M. Boukabous and M. Azizi, “Image and video-based crime prediction using object detection and deep learning,” Bulletin of
Electrical Engineering and Informatics, vol. 12, no. 3, pp. 1630–1638, 2023, doi: 10.11591/eei.v12i3.5157.
[9] S. A. A. Shah, A. H. Emara, A. A. Wahab, N. A. Algeelani, and N. A. Al-Sammarraie, “Street-crimes modelled arms recogni-
tion technique employing deep learning and quantum deep learning,” Indonesian Journal of Electrical Engineering and Computer
Science, vol. 30, no. 1, pp. 528–544, 2023, doi: 10.11591/ijeecs.v30.i1.pp528-544.
[10] G. K. Verma and A. Dhillon, “A handheld gun detection using faster R-CNN deep learning,” in ACM International Conference
Proceeding Series, 2017, pp. 84–88, doi: 10.1145/3154979.3154988.
[11] J. Garcia-Pajuelo and E. Paiva-Peredo, “Comparison and evaluation of yolo models for vehicle detection on bicycle paths,” IAES
International Journal of Artificial Intelligence, vol. 13, no. 3, pp. 3634–3643, 2024, doi: 10.11591/ijai.v13.i3.pp3634-3643.
[12] M. Grega, A. Matiolański, P. Guzik, and M. Leszczuk, “Automated detection of firearms and knives in a CCTV image,” Sensors,
vol. 16, no. 1, 2016, doi: 10.3390/s16010047.
[13] D. Romero and C. Salamea, “Convolutional models for the detection of firearms in surveillance videos,” Applied Sciences, vol. 9,
no. 15, 2019, doi: 10.3390/app9152965.
[14] R. S. Mehsen, “Deep learning algorithm for detecting and analyzing criminal activity,” International Journal of Computing, vol. 22,
no. 2, pp. 248–253, 2023, doi: 10.47839/ijc.22.2.3095.
[15] A. R. Raju, T. Maddileti, S. J, R. Srinivas, and K. Saikumar, “Pseudo trained yolo R CNN model for weapon detection with a
real-time kaggle dataset,” International Journal of Integrated Engineering, vol. 14, no. 7, 2022, doi: 10.30880/ijie.2022.14.07.011.
[16] S. K. Nanda, D. Ghai, P. Ingole, and S. Pande, “Analysis of video forensics system for detection of gun, mask, and anomaly using
soft computing techniques,” AIP Conference Proceedings, vol. 2800, no. 1, 2023, doi: 10.1063/5.0162900.
[17] S. A. A. Akash, R. S. S. Moorthy, K. Esha, and N. Nathiya, “Human violence detection using deep learning techniques,” Journal of
Physics: Conference Series, vol. 2318, no. 1, 2022, doi: 10.1088/1742-6596/2318/1/012003.
[18] H. Gao, “A yolo-based violence detection method in iot surveillance systems,” International Journal of Advanced Computer Science
and Applications, vol. 14, no. 8, pp. 143–149, 2023, doi: 10.14569/IJACSA.2023.0140817.
[19] M. Zahrawi and K. Shaalan, “Improving video surveillance systems in banks using deep learning techniques,” Scientific Reports,
vol. 13, no. 1, 2023, doi: 10.1038/s41598-023-35190-9.
[20] Y. Al-Smadi et al., “Early wildfire smoke detection using different yolo models,” Machines, vol. 11, no. 2, 2023, doi: 10.3390/ma-
chines11020246.
[21] D. A. Cadillo-Laurentt and E. A. Paiva-Peredo, “Histopathological image classification using convolutional neural networks for
detection of metastatic breast cancer in lymph nodes,” International journal of online and biomedical engineering, vol. 20, no. 2,
pp. 31–45, 2024, doi: 10.3991/ijoe.v20i02.46789.
[22] V. E. Sathishkumar, J. Cho, M. Subramanian, and O. S. Naren, “Forest fire and smoke detection using deep learning-based learning
without forgetting,” Fire Ecology, vol. 19, no. 1, 2023, doi: 10.1186/s42408-022-00165-0.
[23] F. M. Talaat and H. ZainEldin, “An improved fire detection approach based on yolo-v8 for smart cities,” Neural Computing and
Applications, vol. 35, no. 28, pp. 20939–20954, 2023, doi: 10.1007/s00521-023-08809-1.
[24] P. Mehta, A. Kumar, and S. Bhattacharjee, “Fire and gun violence based anomaly detection system using deep neural networks,”
in 2020 International Conference on Electronics and Sustainable Communication Systems (ICESC), IEEE, 2020, pp. 199–204, doi:
10.1109/ICESC48915.2020.9155625.
[25] L. Zhao, L. Zhi, C. Zhao, and W. Zheng, “Fire-YOLO: a small target object detection method for fire inspection,” Sustainability,
vol. 14, no. 9, 2022, doi: 10.3390/su14094930.
[26] S. N. Saydirasulovich, A. Abdusalomov, M. K. Jamil, R. Nasimov, D. Kozhamzharova, and Y.-I. Cho, “A YOLOv6-based improved
fire detection approach for smart city environments,” Sensors, vol. 23, no. 6, 2023, doi: 10.3390/s23063161.
[27] H. Zheng, J. Duan, Y. Dong, and Y. Liu, “Real-time fire detection algorithms running on small embedded devices based on Mo-
bileNetv3 and YOLOv4,” Fire Ecology, vol. 19, no. 1, 2023, doi: 10.1186/s42408-023-00189-0.
[28] J. Lin, H. Lin, and F. Wang, “A semi-supervised method for real-time forest fire detection algorithm based on adaptively spatial
feature fusion,” Forests, vol. 14, no. 2, 2023, doi: 10.3390/f14020361.

BIOGRAPHIES OF AUTHORS

Ederson Rodriguez-Rosas is student as an electronic engineer from the Universidad

Tecnológica del Perú. His learning interests include industrial process automation, machine learning,
and neural networks. He currently holds the position of intern in the TEC operations area, which is
focused on the management of technological projects in the field of electronic security, which consist
of the implementation of CCTV systems, access control, fire and intrusion detection. He can be
contacted at email: [email protected].

Int J Artif Intell, Vol. 14, No. 2, April 2025: 945–953

Int J Artif Intell ISSN: 2252-8938 ❒ 953

Aron Castillo-Turpo is student as an electronic engineer from the Universidad Tecnológica

del Perú. His learning interests include data communication networks, as well as control and automa-
tion. He currently holds the position of supervisor in the telecommunications area at Claro, which is
focused on quality control and large-scale fiber optic projects, which consists of offering and provid-
ing telecommunications services with the highest quality. Coverage and innovation that meet people’s
needs. He can be contacted at email: [email protected].

Kevin Acuña-Condori received his B.Eng. degree in electrical engineering from Univer-
sidad Nacional Mayor de San Marcos, Perú, in 2015. He was awarded the CONCYTEC scholarship
in 2015 for his master’s studies, which he completed in mechatronic engineering from the Pontif-
ical Catholic University of Peru, Lima, in 2017. Currently, he serves as a lecturer in electronic
design, digital systems, control and automation, digital image processing, artificial intelligence, and
robotics at the Universidad Tecnologica del Perú since 2017 and Pontifical Catholic University of
Peru since 2018. As a researcher at Universidad Tecnológica del Peru, he has authored publications
in bioengineering, artificial intelligence, and control. His research interests include bioengineering,
artificial intelligence, brain-computer interface, and neuroprosthetics. He can be contacted at email:
[email protected].

Ernesto Paiva-Peredo is graduated as an electrical mechanical engineer from the Univer-

sidad de Piura in Peru in 2013. He pursued further studies and obtained a master’s degree in electrical
mechanical engineering with a specialization in automation and optimization, which was supported
by CONCYTEC and awarded by the Universidad de Piura in 2016. During his academic journey,
he served as a research assistant at the Department of Technology and Innovation (DTI) at SUPSI.
Presently, he holds the position of a Professor-Researcher at the Universidad Tecnológica del Perú.
He can be contacted at email: [email protected].

Security in smart cities using YOLOv8 to detect lethal weapons (Ederson Rodriguez-Rosas)