3D Vision Object Identification Using YOLOv8

3D VISION OBJECT IDENTIFICATION USING YOLOv8

Miguel Silveira1, Adriano A. Santos 1,2, Filipe Pereira 1,2,*, António Ferreira da Silva1,2,
Carlos Felgueiras3, António Ramos2 and José Machado4
1CIDEM—School of Engineering (ISEP), Polytechnic of Porto (P.Porto),

4249-015 Porto, Portugal;

2INEGI—Institute of Science and Innovation in Mechanical and Industrial Engineering,

4200-465 Porto, Portugal

3CIETI—Centre of Innovation on Engineering and Industrial Technology (ISEP),

School of Engineering, 4249-015 Porto, Portugal

4MEtRICs—Research Centre, School of Engineering, University of Minho,

Campus of Azurém, 4800-058 Guimarães, Portugal

Email: fal@isep.ipp.pt

Abstract - This article aims to explore the application of 3D vision techniques in the automation of
logistical processes, using deep learning for object identification and manipulation in industrial
environments. The work develops a 3D vision system that employs object and keypoint detection
models trained with tools such as Roboflow and YOLOv8 (You Only Look Once version 8). The
methodology includes data collection and annotation, development of deep learning models, and
analysis of the obtained results. The models demonstrated high precision and recall in the block
and keypoint identification, with a slight reduction in keypoint model accuracy due to variability in
annotations. Integrating the models presented computational challenges, but the combined
approach proved effective in precise detection. Limitations include the need for resource
optimization and improvement in annotation processes. In addition, the accuracy obtained is due to
the fact that the object detection system was trained with large amounts of data to provide this high
accuracy. The model was trained and evaluated based on measures of average precision (mAP) and
recovery, obtaining a mAP of 98.3%, a precision of 96.4% and a recall of 95.6%.

Keywords: 3D Vision; Logistics Automation; Roboflow; Object Detection; Keypoints; YOLOv8.

1. Introduction their effectiveness in object detection and

recognition [3].
The automation of logistical processes has garnered Another significant study, "Pick and Place Robotic
significant attention due to the pressing need to Arm: A Review Paper," provides a comprehensive
enhance efficiency and reduce operational costs. review of pick-and-place robotic arms, discussing
Advances in technology, particularly in the realms of control technologies, feedback sensors, and
computer vision and artificial intelligence, have integration with computer vision systems. The
enabled the development of more intelligent and combination of these technologies has been shown
adaptive systems [1]. Among these, 3D vision stands to enhance the accuracy, speed, and adaptability of
out for its substantial potential to improve the robotic systems. Despite these advancements,
precision and flexibility of automated systems, several gaps remain, notably the computational
allowing for more effective detection of objects. This complexity involved in integrating multiple
technological evolution has driven research and computer vision models. Most previous works have
development towards innovative solutions that can focused on either object detection or manipulation,
be integrated into industrial processes, offering with few exploring the simultaneous integration of
considerable gains in productivity and quality [2]. object detection and keypoint detection models for a
A notable example within this technological more complete solution [4].
advancement is the "Amazon Picking Challenge," The primary objectives of this work are to
where teams devised robotic solutions for pick-and- develop a 3D vision system capable of identifying
place tasks in warehouses. These teams faced objects in a logistical environment using state-of-the-
challenges such as handling objects of varying art deep learning technologies. This involves
shapes, sizes, and materials and integrating precise implementing and training a set of object detection
and efficient artificial vision systems. Techniques and keypoint detection models using tools such as
like YOLO (You Only Look Once) were highlighted for Roboflow and YOLOv8. A significant aspect of this

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 7

 3D Vision Object Identification Using YOLOv8

work is the merging of these two vision models to The parallelepipeds used in the simulation can be
create a comprehensive solution for logistical seen at Figure 1.
automation, including the detection of the center of
mass of objects, which is crucial for tasks involving
manipulation, ensuring stability and accuracy during
handling. This study contributes to the field by
addressing the computational challenges associated
with integrating 3D vision models and
demonstrating their practical application in
automating logistical processes. The findings suggest Figure 1: Parallelepipeds used for YOLO identification
that while the integrated approach is effective, there
are limitations related to resource optimization and
annotation processes that need further refinement. 2.1 Tools and Technologies
The principal conclusions highlight that the i. Software Tools
combined approach of using object detection and
keypoint detection models is highly effective, The open-source software Python was chosen for
achieving a mean Average Precision (mAP) of 99.0% its versatility and extensive library support for
for block identification and 98.3% for keypoint machine learning and computer vision, including
detection. The detection of the center of mass further Numpy, TensorFlow, and PyTorch. These libraries
enhances the system's capability, indicating its facilitated the rapid development of complex
potential for significant advancements in industrial algorithms necessary for this project [9][10]. To
automation, especially in applications requiring support development, a cloud-based platform,
precise manipulation of objects. Roboflow, was used to perform all the data
annotation and train the object detection models.
2. Materials and Methods This platform simplifies image pre-processing,
normalisation and the creation of training and
validation datasets, automating significant parts of
In this investigation, we used a version of the YOLO
the data preparation process [11]. The YOLOv8 (You
algorithm: YOLOv8 [5]. YOLOv8 is a fast, high-
Only Look Once) was selected for its high precision
performance object detection algorithm used for
and speed in detecting objects in real time. This
real-time object detection. Objects, when integrated
model processes entire images in a single pass,
into an industrial process, can move at high speeds,
which makes it ideal for applications that require
so the detection speed must also be high. YOLOv8 is
fast and accurate detection [12]. YOLO makes use of
able to fulfil this requirement. The algorithm was
1x1 convolutions to define object classes, delimiting
developed using Roboflow and YOLOv8, an open-
them with boxes, and assigning class probability
source deep learning framework, which made it easy
scores [13].
to train and test customised data sets and offers
excellent detection performance even in extreme
situations, such as detecting potholes in a road [6],
drones [7] and others. In order to achieve these
research objectives, a series of steps were
implemented, which are described below:
• Identifying objects that could be used to
simulate an industrial environment.
• Acquisition of the images needed to train,
validate and test the model from a 70:30
perspective.
• Annotating the data by labelling it with the
information needed to train the models. Figure 2: Intel RealSense Depth D435i camera
• Identification of keypoints in the model's data
set. ii. Hardware Tools
• Training model and identification of results
To simulate a typical industrial process, five Two types of cameras were used: a Xiaomi Note
wooden parallelepipeds were used, each measuring 12 smartphone camera (50 MP resolution, f/1.8
75 mm x 25 mm x 15 mm. These blocks, uniformly aperture) for initial data collection and the Intel
colored in yellow, blue, orange, green, and purple. RealSense D435i camera (Figure 2) for final
These were used to simulate the loads in a implementation. The Intel RealSense D435i includes
palletising simulation, a common logistics task [8]. depth sensors and an inertial measurement unit
Their uniform weight distribution ensured that their (IMU) for precise depth data and six degrees of
geometric centre coincided with their centre of mass. freedom (6DoF) measurements.

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 8

 3D Vision Object Identification Using YOLOv8

2.2 Data Collection

Images of the blocks were captured in a

controlled environment to ensure consistency and
reliability of the dataset, as shown in Figure 3. The
images were taken under different lighting
conditions and from multiple angles to ensure a
diverse and comprehensive dataset. This diversity
helps the model generalize better to real-world
scenarios where lighting and object positions can
vary significantly.

Figure 5: Example of an image with the blocks and

other items

2.3 Data Annotation

After collecting the images, the next step was data

annotation, where each image was labeled to include
the necessary information to train the models. The
collected images were uploaded to the Roboflow
platform, which offers intuitive tools for data
annotation, as shown in Figure 6.

Figure 3: Controlled environment for data collection

The data collection process involved capturing a

total of 200 images, divided into two sets:
• 100 images with only the five blocks in various
arrangements to capture different perspectives
and orientations. These images provide clean
and clear data, allowing the model to learn the
characteristics of parallelepipeds without Figure 6: Roboflow annotation display
interference from other objects, as shown in the
Figure 4. Each parallelepiped in the images was manually
annotated, and the classes created for each model
were as follows: Block Identification Model: “Yellow
Block”, “Blue Block”, “Orange Block”, “Purple Block”,
“Green Block”; Keypoint Identification Model:
“Yellow-Block-KP”, “Blue-Block-KP”, “Orange-Block-
KP”, “Purple-Block-KP”, “Green-Block-KP”.

Figure 4: Example of an image with blocks only

• 100 images with additional objects placed

around the blocks to simulate a more realistic
and cluttered logistical environment, as shown
in the Figure 5. Figure 7: “Blue-Block-KP” Keypoint Annotation
Structure

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 9

 3D Vision Object Identification Using YOLOv8

For the keypoint identification model, each class essential for replicating and building upon the
had keypoints annotated in a structure as illustrated results presented in this research.
in Figure 7, with the following vertices: “Vertice-BA”,
“Vertice-BAZ”, “Vertice-BL”, “Vertice-BR”, “Vertice- 2.4 Model Trainng
BV”. 2.4.1 Experiment and Data Acquisition
For the object identification model, the bounding
boxes were tightly drawn around each As already mentioned, the images were taken
parallelepiped, as shown in Figure 8. with two different cameras. The images of the
objects were taken from different angles and views,
ensuring the variability of the data set.
The dataset consisted of 200 annotated images
and was divided into sets, with 70% used for
training, 20% for validation and 10% for the final
test, following Roboflow's recommendations, 70:30
train:test. This division ensured a balanced approach
to training, validating and evaluating the model. This
data splitting approach was applied to both models,
the block identification 158 model and the keypoint
identification model, ensuring consistency in the
training, validation, and testing processes. So, data
preprocessing steps included auto-orientation and
resizing images to 640x640 pixels. Data
augmentation techniques, such as random flipping
and brightness adjustments, were applied to
Figure 8: Annotated image of the model dataset – enhance model robustness against different
“Block Identification” orientations and lighting conditions.
Roboflow automatically adds two steps to the
For the keypoint identification model, the preprocessing: auto-orientation and resizing to
annotation boxes did not allow for the precise 640x640. Auto-orientation ensures that the images
contouring of the piece, as only rectangles could be are stored on disk in the same way that applications
drawn [14][15]. This simplifies the annotation open them. Automatic orientation ensures that
process and focuses on the key points (vertices), as images match the orientation of the source device
shown in the Figure 9. and prevents incorrect data from being entered into
YOLOv8. Resizing or trimming creates a consistent
size for the images, facilitating and speeding up the
training process. For both models, auto-orientation
and resizing were chosen as preprocessing steps. In
Green-Block-KP addition to pre-processing the data, we utilized the
Roboflow API, which helped mitigate the issue of
missing data. We set the following parameters:
inversion: horizontal and vertical, tonality: between -
25 degrees and +25 degrees, and cropping.

Figure 9: Annotated image of the model dataset –

“KeyPoints Identification”

The annotated datasets for both object detection

and keypoint detection models have been deposited
in a publicly available database. These datasets can
be accessed on the Roboflow Universe website by
searching for the respective model names. The
datasets include detailed annotations and are Figure 10: Augmented dataset

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 10

 3D Vision Object Identification Using YOLOv8

Data augmentation ensured data variability by The block identification code utilized YOLOv8 to
artificially generating images. For the object detect and annotate blocks in real-time. The
detection model, random flip augmentation was processing pipeline involved capturing video frames,
applied both horizontally and vertically. This running inference, and displaying annotated frames
technique increases data variability and helps the with confidence levels and labels. Therefore, the
model identify blocks regardless of their orientation, keypoint identification code was developed to detect
enhancing the model's ability to generalize from the vertices of the blocks, which are essential for
different perspectives. Figure 10 shows augmented calculating the center of mass. The process involved
dataset and Figure 11 shows flowchart experiment. directly marking visible vertices and estimating
positions for occluded ones. Then, the object
detection and keypoint detection models were
integrated to create a cohesive system. This
integration enabled precise and robust identification
of objects and their key features, facilitating accurate
logistical processes. The algorithm used can be
implemented using the pseudocode presented in
Algorithm 1.

Figure 11: Flowchart experiment Algorithm 1

Import libraries
On the other hand, for the keypoint detection # create annotation
model, brightness modification (±15%) was applied Annotator = sv.BoxAnnotator()
to improve model robustness against different # defineshared state
shared_state = {keypoints and object predictions}
lighting conditions. This augmentation ensures that
def custom_sink_keypoints(predictions):
keypoints are detected accurately even under | image = image_frame.image.copy()
varying illumination, enhancing the model's | # calculete center of mass
performance in diverse lighting scenarios. | for prediction in predictions:
| | if “keypoints” in prediction:
2.5 Training Block and Keypoint | | | keypoints = prediction
Identification | | | keypoint_coords = np.array [x,y]
| | | center_of_mass = np.mean ()
The models were trained using YOLOv8 on the | return
def custom_sink_objects(prediction):
Roboflow platform, which automates the training | labelling objects
environment setup, including hardware and | image = box_annotator.annotator()
software configurations, thus reducing the time | # display annotated images
needed to start training. | cv2.imshow(“combined annotations”, image)

Figure 12: Roboflow Training Metrics for Object Identification Model

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 11

 3D Vision Object Identification Using YOLOv8

3. Results The precision metric (metrics/precision(B))

evaluates the proportion of true positives relative to
3.1 Block Identification Results the total predictions, while recall (metrics/recall(B))
measures the model's ability to identify all positive
The model trained was evaluated using for mAP,
examples. Both curves are high, indicating the
precision, and recall. Model was tested on a dataset model's high accuracy and identification capability.
generated from a randomized training split. The test The "val/box_loss" and "val/cls_loss" measure the
images showed variability in the data due to the use losses in bounding box predictions and classification
of different backgrounds. Based on this, the block on validation data, showing that the model
identification model showed the following results: a generalizes well and maintains accuracy on unseen
mAP of 99.0% with a Precision of 97.9% and a Recall data. The "val/dfl_loss" confirms the improvement in
of 96.4%. Figure 12 provide a detailed view of the the confidence of predictions on validation data. The
"metrics/mAP50(B)" evaluates the mean average
performance of the object detection model.
precision of the model's predictions with an
The "train/box_loss" metric refers to the loss in intersection over union (IoU) of 50%, indicating
predicting bounding boxes during training, good performance in object detection, while the
indicating a constant improvement in the accuracy of "metrics/mAP50-95(B)" evaluates the average
the predicted boxes compared to the actual boxes. precision at different IoU levels, suggesting that the
The "train/cls_loss" measures the classification loss, model is robust and accurate across various overlap
showing that the model is becoming more accurate criteria.
in classifying objects within the bounding boxes. The
"train/dfl_loss" adjusts the loss based on the 3.2 Keypoint Detection Results
confidence distribution of the predictions, indicating
The keypoint detection model shows the
that the model is learning to make more reliable following results: a mAP of 98.3% with a Precision of
predictions. 96.4% and a Recall of 95.6%.

Figure 13: Roboflow Training Metrics for Keypoints Detection Model

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 12

 3D Vision Object Identification Using YOLOv8

The metrics presented in Figure 13 provide a of mass were determined using pixel coordinates,
detailed view of the performance of the keypoint and the bounding boxes and class score.
detection model. The "train/dfl_loss" metric refers to Table 1 and summarize the means and standard
the "distribution focal loss" during training, deviations of the main metrics. To calculate the
adjusting the loss based on the confidence percentage error, the distance between the
distribution of the model's predictions and calculated center of mass and the actual center of
penalizing high-confidence incorrect predictions. mass was divided by the diagonal distance of a block.
The curve of this metric shows a constant decrease,
indicating that the model is learning to make more Table 1: Center of Mass Calculation Error per Block
reliable predictions. Color and Standard Deviation
The precision metrics (metrics/precision(B) and Average Standard
metrics/precision(P)) and recall metrics Block
[%] Deviation [%]
(metrics/recall(B) and metrics/recall(P)) evaluate 2.50 1.43
Yellow
the accuracy and the model's ability to correctly
identify keypoints and objects. The curves of these Blue 2.61 1.62
metrics show that the model has high precision and Orange 2.45 1.29
identification capability, with few incorrect Purple 3.28 2.00
predictions. The "val/dfl_loss," similar to Green 2.34 1.33
"train/dfl_loss" but evaluated on validation data, Total error 2.64 1.48
confirms the improvement in prediction confidence
on unseen data during training.
4. Discussion
The "metrics/mAP50(B)" and "metrics/mAP50-
95(B)" metrics evaluate the average precision of the
The main objective of this study is centered on
model's keypoint predictions at different IoU levels,
developing and evaluating models capable of
showing consistent and robust performance. The identifying blocks and keypoints using advanced
"metrics/mAP50(P)" and "metrics/mAP50-95(P)" deep learning techniques. These models, and
metrics do the same for objects, indicating efficiency essentially the determination of keypoints, prove to
in detection. The decreasing loss curves for training be fundamental for robotic manipulation not only in
and validation data suggest that the model is identification, but also in defining the object's
improving its performance over time. The high gripping point to be applied, for example, in [16][17].
precision, recall, mAP50, and mAP50-95 metrics for The results indicate that both models, trained with
keypoints and objects indicate that the model is Roboflow and using YOLOv8, achieved high levels of
efficient and accurate in detection, with good precision and recall. However, there were notable
generalization to new data. differences in the performance of the block
identification model compared to the keypoint
Orange Block (96.8%) detection model.

4.1 Comparison of Center of Mass

Yellow Block (96.1%) Calculation
Blue Block (96.0%)

The accuracy of the center of mass (CoM)

Purple Block (95.4%)
calculations for different colored blocks, a critical
Green Block (97.0%) aspect of this study, showed varying percentage
errors and standard deviations. The yellow block
had a mean percentage error of 2.50% with a
standard deviation of 1.43%, indicating consistent
precision under stable conditions. However, blue
Figure 14: Manual Annotation of the Center Mass block's mean error slightly increased to 2.61% with
a standard deviation of 1.62%, reflecting greater
3.2 Accuracy of the Center of Mass variability. On the other hand, the orange block
Calculation exhibited the lowest mean error at 2.45% and a
standard deviation of 1.29%, demonstrating the
To evaluate the accuracy of the center of mass highest precision and consistency. Conversely, the
calculation for the parallelepipeds, 50 photos were purple block had the highest mean error at 3.28%
taken using the Intel RealSense D435i camera. The and the largest standard deviation of 2.00%,
images were processed with the developed code to indicating the least precision and most variability.
calculate the center of mass of the objects. After Otherwise, the green block, with a mean error of
collecting the data, the actual centers of mass were 2.34% and a standard deviation of 1.33%, also
manually annotated, as shown in Figure 14, and the showed good precision with relatively low
distances between the calculated and actual centers variability.

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 13

 3D Vision Object Identification Using YOLOv8

These findings suggest that while the overall CoM “Online yarn hairiness– Loop & protruding fibers
calculation performance was satisfactory, precision dataset” [22] have shown the potential of algorithms
varied significantly between different colored blocks, like YOLO for improving logistical operations. The
with the high variability in the purple block augmentation techniques guaranteed the robustness
potentially due to specific lighting conditions or of the model, even with variations in brightness,
inherent color characteristics that significantly affect when other objects were included and on the
detection. different backgrounds used. On the other hand, the
results obtained in our study are in line with the
4.2 Analysis of Model Performance values obtained by various scholars in terms of static
identification, which is promising (accuracy of
The identification of blocks proved to be more 96.4%) for future developments.
straightforward compared to keypoint detection due However, this study extends previous work by
to several factors. First, identifying entire objects is integrating both object detection and keypoint
simpler than pinpointing specific points within an detection, providing a more comprehensive solution
object, as the latter requires a detailed for automation. The computational challenges
understanding of the object's internal structure, encountered highlight the need for further research
thereby increasing task complexity and data into optimizing these combined approaches.
variability. Second, annotations for entire objects are Future research should focus on optimizing
generally more consistent because defining and computational resources to enhance the efficiency of
annotating an object's location and shape is more combined models, thereby reducing processing
straightforward, whereas keypoint annotations are times and resource requirements. Additionally, there
prone to greater error and inconsistency, is a need to improve annotation processes by
introducing significant noise into the training data. developing automated or semi-automated tools to
Lastly, object detection models are less increase the consistency and accuracy of keypoint
computationally intensive, requiring fewer layers annotations. Broadening application scenarios is also
and parameters than keypoint detection models, crucial, this involves testing the models in varied
which need deeper and more complex neural industrial contexts to understand their performance
networks to capture fine details, leading to higher under different conditions and with different object
computational complexity and resource demands. types. Furthermore, future work should explore the
Additionally, the study revealed that combining the detection of object orientation, which could be
two models in a single code base posed significant particularly beneficial for robotic arms in logistics
computational challenges, requiring substantial and manufacturing. By addressing these areas,
processing power and resulting in extended future studies can build on the findings of this
execution times and possible interruptions due to research to develop more robust and efficient
high demands on processing and memory resources. automated systems for logistical and manufacturing
Despite these small variations, the results processes.
obtained are in line with various studies carried out
with YOLOv8, both for static objects and for moving Acknowledgements
objects. Thus, Xiao et al. [14] obtained an accuracy
rate of 99.5% when identifying fruit, Wu et al. [15] This work was partially financially supported by
obtained an accuracy of 94.9% when remotely FCT—Fundação para a Ciência e Tecnologia
identifying static objects. Rajesh et al. [18] in the (Portugal), who partially financially supported this
identification of harvestable tomatoes, colour work through the RD Units Project Scope:
perception, the accuracy was 97.23%. On the other UIDP/04077/2020 and UIDB/04077/2020. The
hand, Telaumbanua et al. [19] when applying authors are grateful, too, to the partial financial
YOLOv8 to mobile objects, Vehicle Detection, support provided by Base Funding—
obtained an accuracy of 77% for training and 96% UIDB/50022/2020 (LAETA) of INEGI—Institute of
for testing. On the other hand, Moussaoui et al. [20], Science and Innovation in Mechanical and Industrial
in the automatic identification of vehicle license Engineering, Portugal.
plates, obtained an accuracy of 99% in detection and
98% in character recognition. Luo et. al [21]
obtained a detection accuracy of 95.6% when References
detecting objects transported on conveyor belts.
[1] Russell, S. J., Russell, S. & Norvig, P. Artificial
5. Conclusions intelligence: a modern approach, 4th ed.;
Pearson, 2021.
The results align with previous studies that have [2] ElMaraghy, H., Monostori, L., Schuh, G., &
demonstrated the effectiveness of deep learning ElMaraghy, W. Evolution and future of
techniques in object detection and identification. manufacturing systems. CIRP Annals 2021, 70(2),
Studies such as the "Amazon Picking Challenge", 635-658.
research on "Pick and Place Robotic Arms" [22] and https://doi.org/10.1016/j.cirp.2021.05.008.

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 14

 3D Vision Object Identification Using YOLOv8

[3] N. Correll et al. Analysis and observations from [15] Wu, T., & Dong, Y. YOLO-SE: Improved YOLOv8
the first amazon picking challenge. EEE for remote sensing object detection and
Transactions on Automation Science and recognition. Applied Sciences 2023, 13(24),
Engineering, 15, 489(1), 172–188, 2016. 12977.
https://doi.org/10.1109/TASE.2016.2600527. https://doi.org/10.3390/app132412977.
[4] S. Surati, S. Hedaoo, T. Rotti, V. Ahuja, and N. [16] Santos, A.A., Haladus, J., Pereira, F., Felgueiras,
Patel. Pick and place robotic arm: a review paper. C., Fazenda, R.. Simulation Case Study for
Int. Res. J. Eng. Technol, 8(2), 2121–2129, 2021. Improving Painting Tires Process Using the
[5] R. Bawankule, V. Gaikwad, I. Kulkarni, S. Kulkarni, Fanuc Roboguide Software. In: Silva, F.J.G.,
A. Jadhav and N. Ranjan. Visual Detection of Pereira, A.B., Campilho, R.D.S.G. (eds) Flexible
Waste using YOLOv8. International Conference Automation and Intelligent Manufacturing:
on Sustainable Computing and Smart Systems Establishing Bridges for More Sustainable
(ICSCSS), Coimbatore, India, 2023, 869-873. Manufacturing Systems. FAIM 2023. Lecture
https://doi.org/10.1109/ICSCSS57650.2023.101 Notes in Mechanical Engineering. Springer,
69688. Cham, 2023. https://doi.org/10.1007/978-3-
[6] Saluky, Y. Marine, A. Zaeni, A. Yuliati, O. R. Riyanto 031-38241-3_58.
and N. Bahiyah. Pothole Detection on Urban [17] Alberto Monteiro, Filipe Pereira, Adriano A.
Roads Using YOLOv8. 10th International Santos, José Machado, Miguel Oliveira.
Conference on ICT for Smart Society (ICISS), Modelling and Simulation of a Pick&Place
Bandung, Indonesia, 2023, 1-6. System using Modelica Modelling Language and
https://doi.org/10.1109/ICISS59129.2023.1029 an Inverse Kinematics Approach. International
1192. Journal of Mechatronics and Applied Mechanics
[7] Zhai, X.; Huang, Z.; Li, T.; Liu, H.; Wang, S. YOLO- 2024, 16.
Drone: An Optimized YOLOv8 Network for Tiny dx.doi.org/10.17683/ijomam/issue16.1.
UAV Object Detection. Electronics 2023, 12, 3664. [18] A. Rajesh, S. Muppala and A. Jeyasekar,
https://doi.org/10.3390/electronics12173664. "Identification of Harvestable Tomatoes using
[8] Carata, S.V., Ghenescu, M., Mihaescu R. Real-Time Yolov8. In 3rd International Conference on
Detection of Unrecognized Objects in Logistics Applied Artificial Intelligence and Computing
Warehouses Using Semantic Segmentation. (ICAAIC), Salem, India, 2024, 1258-1263.
Mathematics 2023, 11, 2445. 10.1109/ICAAIC60222.2024.10575283.
https://doi.org/10.3390/math11112445. [19] Telaumbanua, A. P. H., Larosa, T. P., Pratama, P.
[9] S. Sultonov. IMPORTANCE OF PYTHON D., Fauza, R. U. H., & Husein, A. M. Vehicle
PROGRAMMING LANGUAGE IN MACHINE detection and identification using computer
LEARNING. Int. Bull. Eng. Technol. 2023,3(9), vision technology with the utilization of the
28-30. yolov8 deep learning method. Sinkron: jurnal
https://internationalbulletins.com/intjour/index dan penelitian teknik informatika 2023, 7(4),
.php/ibet/article/view/1020. 2150-2157. 10.33395/sinkron.v8i4.12787.
[10] Z. DeVito, J. Ansel, W. Constable, M. Suo, A. [20] Moussaoui, H., Akkad, N.E., Benslimane, M. et al.
Zhang, and K. Hazelwood. Using python for Enhancing automated vehicle identification by
model inference in deep learning. arXiv Prepr. integrating YOLO v8 and OCR techniques for
arXiv2104.00254, 2021. high-precision license plate detection and
https://doi.org/10.48550/arXiv.2104.00254 recognition. Sci Rep 2024, 14, 14389.
[11] S. Alexandrova, Z. Tatlock, and M. Cakmak. https://doi.org/10.1038/s41598-024-65272-1.
RoboFlow: A flow-based visual programming [21] Luo B, Kou Z, Han C, Wu J. A. “Hardware-
language for mobile manipulation tasks. In IEEE Friendly” Foreign Object Identification Method
International Conference on Robotics and for Belt Conveyors Based on Improved YOLOv8.
Automation (ICRA), IEEE, 2015, 5537–5544. Applied Sciences 2023; 13(20):11464.
https://doi.org/10.1109/ICRA.2015.7139973. https://doi.org/10.3390/app132011464.
[12] M. Hussain. YOLO-v1 to YOLO-v8, the rise of [22] Santos, A.A., Pereira, F. & Felgueiras, C.
YOLO and its complementary nature toward Optimization and improving of the production
digital manufacturing and industrial defect capacity of a flexible tyre painting cell. Int J Adv
detection. Machines 2023, 11(7), 677. Manuf Technol 2024.
https://doi.org/10.3390/machines11070677. https://doi.org/10.1007/s00170-024-13208-4.
[13] Aydin, B.; Singha, S. Drone Detection Using [23] Filipe Pereira, Leandro Pinto, Filomena Soares,
YOLOv5. Eng 2023, 4, 416-433. Rosa Vasconcelos, José Machado, Vítor
https://doi.org/10.3390/eng4010025. Carvalho, Online yarn hairiness– Loop &
[14] Xiao, B., Nguyen, M. & Yan, W.Q. Fruit ripeness protruding fibers dataset, Data in Brief, Volume
identification using YOLOv8 model. Multimed 54, 2024, 110355, ISSN 2352-3409,
Tools Appl 2024, 83, 28039–28056. https://doi.org/10.1016/j.dib.2024.110355.
https://doi.org/10.1007/s11042-023-16570-9.

International Journal of Mechatronics and Applied Mechanics, 2024, Issue 17 15