SHODH SAGAR®

Darpan International Research Analysis

ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed

DETECTION ALGORITHM FOR DETECTING DRONES/UAVs

Armaan Parreka,1,∗ Vanshika Singh1,2,

aElectronics and Communication Engineering aElectronics and Communication Engineering ,
, IIT Roorkee, Uttarakhand, India IIT Roorkee, Uttarakhand, India
[email protected]

Akshat Jain1,3
aElectronics and Communication
Engineering , IIT Roorkee, Uttarakhand,
India
[email protected]

DOI: http://doi.org/10.36676/dira.v12.i3.93

Accepted : 10/09/2024 Published 13/09/2024 * Corresponding author

Abstract
Unmanned Aerial Vehicles (UAVs), also popularly known as drones, have had an exponential evolution
in recent times. This has resulted in better and affordable artifacts with applications in numerous fields.
However, drones have also been used in terrorist acts, privacy violations and involuntary accidents
in high risk zones. To address this problem, for our final year project we are working on studying and
implementing various techniques and algorithms to automatically detect, identify and track small drones.
We did a literature survey on the current deployed methodologies. Many state of the art techniques in
recent times include Radio-Frequency, Audio-based and Radio-Frequency based methods. We mainly
focused on video surveillance methods supported by computer vision algorithms. We used YOLOv5
architecture and implemented background subtraction methods within it. We modified the network to
incorporate these methods. We further tested our model with the test dataset and compared the results
with the benchmark models. We compared our results with the state of the art models based on visual
data. We deployed our model to identify and locate drones and birds using a live camera in real time.
We also tested a pruned version of our model to further improve the result. Further, we examined all
the possible improvements and modifications that could be applied to our existing model to enhance the
evaluation metrics.
Keywords: Unmanned Aerial Vehicles, drones, Radio-Frequency, YOLOv5

1. Introduction
Small and remotely controlled unmanned aerial vehicles (UAVs), also called drones, are of great
benefit to society. They have grown in popularity as a result of rapid technological advancements in
both their hardware and software, including the addition of cameras and audio recording technology, as
well as the support of autonomous flying and human tracking capabilities. Drones are used in a variety of
everyday tasks, including vegetation monitoring, delivery, rescue missions, and security. Despite these
advantages, there has been a fast surge in the usage of drones for bad purposes such as invading

340
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed

privacy, security, and obstructing safety standards. Drone attacks on airports and drug smuggling
usingdrones have both occurred. Drones are used to spy on and record films and audio snippets of people
intheir homes, raising similar privacy issues. Drone Detection is a class of problems to tackle the issue
ofmisuse of drones. It focuses on Detecting, Localizing, Tracking and taking the necessary measures to
control drones. At the moment, several surveillance and detection technologies are being investigated,
each with its own set of tradeoffs in terms of complexity, range, and capability. Radar, radio frequency
(RF), acoustic sensors, and video surveillance with computer vision algorithms are the basic models
that can be employed for drone detection and classification activities.
• Radar: Radar is a classic sensor that allows reliable detection of flying objects at long distances
and near-unaffected performance in adverse lighting and weather conditions. Despite this, they
frequently fail to detect small commercial UAVs with non-ballistic trajectory velocities. Radars
areunable to discriminate between birds and drones due to a lack of precision in distinguishing
the two. They are a costly solution due to their complicated installation and high cost.
• Visual Data: To detect drones, this method employs video or image recognition techniques. Al-
though these methods have shown to be efficient in ideal settings, their performance is highly
influenced by external elements such as weather, dust, fog, or rain, as well as other flying objects
that may resemble drones, such as birds. Aside from their sensitivity, occlusion is another key
challenge.
• Radio-Frequency: One of the most common anti-drone systems on the market is an RF-based
UAV detection system, which detects and classifies drones based on their RF signatures. How-
ever, not all drones use RF transmission, therefore this method isn’t effective for detecting un-
manned aerial vehicles (UAVs) that aren’t connected to the internet.
• Acoustic Sensors: Acoustic detection systems use a network of auditory sensors or microphones
to recognise distinct acoustic patterns of UAV rotors, even in low-light situations. The maximum
operational range of these systems, however, is less than 300 meters. Furthermore, the sensitivity
of these systems to ambient noise, particularly in urban or industrial locations, as well as windy
circumstances, has an impact on detection performance.

Figure 1: Anti Drone System

The above figure 1 shows the working flow of an anti-drone system. The existence of a drone within
a limited area is identified in the first stage. The system then determines if the drone is authorized or
illegal by assessing its features, such as the kind or model of the drone. The system should then be able
to track and locate the drone. In the end, the system obstructs the drone’s objective by employingseveral
traditional mechanisms such as shooting drones with guns, nets, or spoofing and jamming tactics. We
will be working on the Detection and Tracking aspect of our Bachelor’s Degree project usingvisual data
from a static camera.

341
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
2. Moving Object Detection
This section will present the core concepts of image processing and deep learning utilized in the
proposed solution. In section 2.1, the YOLO family of models are discussed. Background subtraction is
a useful technique for object detection problems which we have discussed in section 2.2. Section 2.3
introduces Pruning which is a technique to improve the weights of any deep learning model. In section
2.4, we discuss the Drone vs Bird challenge. Followed by section 2.5 in which we discuss Performance
Evaluation criteria for the proposed solution. Moving Object Detection is an important concept of Image
Processing and Computer Vision. Automated Video surveillance has been used in many sectors these
days. The basic Framework of Video surveillance consists of Environment Modeling, Motion Segmen-
tation, Object Classification and Object Tracking. In Environment Modeling we basically have to recover
and update the Background from the dynamic frames of a video taking into consideration the factors like
sunlight, shadows, moving branches etc. In motion segmentation, one aims to find the regions which
are related to the moving object. Most of the Motion segmentation methods these days use the temporal
or spatial content of the image sequence or frames, the video is broken into. Some famous methods are
Background subtraction, Temporal differencing and optical flow. Then we basically classify the object
on the basis of shape, motion, feature, color or texture. After this we need to track the object from one
frame to the next frame. Given below is a flow chart of the basic flow of moving object detection:

Figure 2: Basic Flowchart of Moving Object Detection In video data [5]

2.1. 2.1 YOLO

YOLO was first introduced in 2015 by Joseph Redmon et al. “You Only Look Once: Unified, Real-
Time Object Detection”. The motivation was to create a unified model of all phases. The system
produces predictive vectors for all objects present in the image after passing through a single network
of multiple CNNs. YOLO avoids iterating over different regions of an image instead it computes all the
features of the images and makes predictions for all the objects. YOLO applies a grid cell on the image
and checks whether the center of the object falls into the grid cell. If yes then that particular grid cell is
responsible for identifying the object even if there are multiple instances of the object in different grids.
Each grid predicts a bounding box based on the confidence score given by EQUATION where p(object)
represents the probability of the object being inside the bounding box and IOUpred truth over the union
of the prediction box and ground truth. YOLO removes all the unnecessary bounding

342
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
boxes which do not contain objects or bounding
boxes which contain objects which are already
present in other bounding boxes using Non
maximum suppression.

Figure 3: Yolov5 Architecture Model [6]

YOLOv5 is the latest version of YOLO models and

it was the first model which was written exclusively for Python. YOLOv5 architecture can be divided into
three parts. First is its backbone which is similar toYOLOv4 and contains its main feature extractor CSP.
Then there is its neck part containing SPP block,PANet etc. The last part is its head part which is derived
straight from YOLOv3.
2.2. 2.2 Background Subtraction
Background Subtraction as the name just suggests are the techniques to remove background from
images. There are different methods to achieve this. Some popular methods are using Gaussian Mixture
models or Bayesian Segmentation. These probabilistic models are used to identify the color and other
background related things in the image to cancel it out.

Figure 4: Background Subtracted Image

from our dataset

2.3 Pruning
2.3.
Pruning is a data compression technique used to
remove unnecessary information which is redun- dant
and non critical for classification. Pruning reduces overfitting by simplifying the final classifier. It
reduces the size of the decision tree without affecting the predictive accuracy of the classifier. Too large
decision trees are vulnerable to overfitting and new instances of data can significantly impact the
accuracy of the model.

Figure 5: Before and After Pruning Result [7]

2.4. 2.4 Drone Vs Bird Challenge

The “International Workshop on Small-Drone Surveillance, Detection and Counteraction Techniques”
(WOSDETC) proposed the Drone Vs Bird Challenge to invite researchers from all around the world to
work on the problem of detecting drone and distinguishing it from birds from a far away distance. It
requires one to create a system that raises alarm when there is a drone in sight. The Organization
provides a dataset which has been regularly updated with each new installment of the challenge. The
shorts video include many different types of drones with a diverse background ranging from mountains,

343
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
sky, buildings etc. The developed algorithms should strive to precisely localize drones and generate
bounding boxes as close to the targets as possible. The Averaged Precision metric (AP) will be used to
assess the results. We mostly used the dataset from the challenge’s 2020 iteration in our project.

2.5. 2.5 Performance Evaluation Criterion

Any classification is evaluated on the basis of number of instances correctly classified and number
of instances miss-classified. True positive refers to those instances which are correctly classified. False
positives are those instances which were misclassified as positive. True negatives are those instances
which are false in both predicted data and true data. Similarly we can say for False negative as well.
These four parameters are used to make all the important metrics considered for evaluation.
TP, TN, FP and FN are true positives, true negatives, false positives and false negatives respec-
tively.
• Accuracy It is a ratio of correct prediction and total number of predictions.
TP + TN
Accuracy =
TP + TN + FP + FN
(1)
• Precision It is a ratio of correctly predicted outcomes over all the positive outcomes.
TP
Precision =
TP + FP (2)
• Recall It measures the sensitivity of the model and is a ratio of correctly identified positive to
allthe positively identified examples.
TP
Recall =
TP + FN (3)
• F1 score It is the harmonic mean of precision and recall. It gives the performance of the model
interms of both precision and recall.
Precision · Recall
F1score = 2
Precision + Recall (4)

344
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed

3. Proposed Framework
3.1. Model - An altered version of YOLOv5
Our main idea was to use YOLOv5 along with background subtraction to improve YOLO capabilities
for tiny objects like drones. Initially, we tried to implement this through two modules. One to identify
drones solely using YOLOv5, then passing the result to a classifier identifying bird and drone. If the
results are positive with a small probability the image would be then passed on to a classifier that uses
background-subtracted images. So, in Module 1 we apply a basic Yolov5 model and if we get a confi-
dence value above a particular threshold (we choose 0.15 here), then we move to a bird-drone classifier
to differentiate between the detected objects. If the confidence value fails to meet the threshold we pass
to module 2. In module 2, we planned to employ Background Subtraction. We start by extracting Back-
ground Subtracted frames, then dilation. In Dilation we add more pixels to the boundaries of objects
in an image.This connects the closely spaced pixel and helps to reduce
the regions to be checkedby the classifier. Then we applied morphological filtering followed by
deploying MobileNetv2, a small lightweight CNN model. Then we again passed the results to a bird
Drone classifier used in Module 1 as well. We tried to implement this initial approach but we couldn’t
find the desired results. We used an ensemble inherited in the YOLOv5 repository forthis purpose. To
our surprise, our ensemble did notimprove the results. On the test dataset, this ensemble was only able
to produce 58.9% which even both models individually were delivering. This was because the
background subtracted image’s features were drastically different from a normal image. So, their
correlation was poor. A major disadvantage of the YOLOv5 model is that it is not adaptive and so we
tried to merge its architecture itself with adaptive background subtraction techniques rather than
combining three different models. So we switched to trying to integrate background subtraction
capability into existing YOLOv5 architecture.

Figure 6: Initially Proposed Model-Module 1

345
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
Figure 7: Initially Proposed
Model-Module 2

YOLOv5 architecture can be broadly

divided into three parts. Each part
consists of different types of CNN with
different parameters. The first part is
called its backbone, which consists of its
main feature extractor, an improved
version of CSP Darknet. The second part
is its neck layer which divides detection
into three parallel streams based on
dense or sparse prediction. The final
layer, also known as the output layer,
consists of a detect node that combines
the output of the three parallel detectors.
Our main focus is to include a
background subtraction method in YOLOv5 architecture to improve the detection of objects which are
very similar to the background and often go undetected. So, to incorporate the background subtraction
method we first used OpenCV methods and a Convolutional Layer to process the image in the
background-subtracted frame. We used a CNN based Classifier to process this imageparallelly with the
YOLOv5 architecture and get the final output. We merged it with the detect node, as mentioned earlier.
In this way, instead of using an ensemble of two models, we have merged the two processes BG
Subtraction and YOLOv5 during the training itself. To add this background subtracted model in the
YOLOv5, we first need to understand the working of the YOLOv5 repository. The most important code
files for YOLOv5 are present in the model folder of the repository. The file common.py contains all the
required classes of different types of neural nets used in its architecture. We added a new class for
background subtracting Conv2D to it. We used OpenCV to run the background subtraction after
converting the PyTorch tensor to NumPy array, running background subtraction on it, then again
converting it back to the PyTorch tensor. The file yolo.py is where all the model generation is present.
We have modified the class model present in yolo.py to add our background-subtracted layer. Yolov5
takes input for all the hyperparameters using yaml files which are also processed in yolo.py only. We
modified it as well to accommodate new hyperparameters.

3.2. Dataset
We prepared the data set for training the YOLOv5 model using the Drone vs Bird 2020 challenge
dataset and random internet videos. The drone vs. bird training set consists of 77 different videos with
annotations. There are 14 more videos in the challenge test set that do not have annotations. These
videos share characteristics with the training set. To use YOLOv5, we first segmented these videos to
obtain images, and then labeled them with a labeller that uses OpenCV and Tracking algorithms to make
labeling a larger set of images easier. The videos comprise 1384 frames on average and 3 different
resolutions, namely 1920 × 1080 @25 fps, 720 × 576 @50 fps, 1280 × 720 @30 fps. We labeled the
images on our own as we needed the normalized values and in the same format in which YOLOv5 takes
the annotations.

346
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed

Figure 8: Different Drones in the Videos of the Dataset [16]

3.3. Dataset for the initial approach

We prepared three datasets for training the YOLOv5 model from videos of resolution 1920 x 1080:
• This is the simplest data set which was made just to check that YOLOv5 can detect far away small
objects too just like it detects normal moving objects like cars at traffic lights etc. We took a single
video and extracted 48 frames from it. We divided the frames into train images of 40 framesand 8
frames of validation images. We manually labeled the images. Then we trained the model using
Google Collab. To improve the generalization of the model we should use more imagesfor
training and validation, frames from video with different backgrounds including sequences with
sky or vegetation as background, different weather types (cloudy, sunny), direct sun glare and
variation in camera characteristics. The validation and training set should also contain a different
set of images( from different sequences) to avoid overfitting of the model.
•

Figure 9: Training Result using the First Dataset

Figure 10: Test Result using the First Dataset

347
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
• To solve the previous problem (wrong detection of the moon as a drone), we will create a data set
of contrasting videos and about 300-400 frames. We used only one class that is ‘Drone’ here as
well. For train images we had 538 frames from different videos and for validation images we had
180 frames. Because the model was only trained to detect one type of flying item (drones),
the detector is biased toward detecting drones and may fail to distinguish them from other similar-
looking flying objects (especially if they are really small). To eliminate this bias, you might either
train for more object classes (acquire and annotate more photos including things like birds, planes,
and helicopters) or use several frames to obtain information about their flying patterns.

Figure 11: Test Result using the Second Dataset

• In the third data set we created two classes of ‘Drone’ and ‘Bird’ and used about 150-180 frames
in the train images folder and 50-80 frames for validation images. We used video of birds to train
the model for different bird motions in the sky.

Figure 12: Test Result using the Third Dataset

3.4. Dataset for the altered approach

For this, we created a Train dataset of about 600 images, a validation dataset of about 200 images,
and a test dataset consisting of 150 images. We planned to train with a larger number of images to
improve the results. In creating this dataset we focused more on making the train and validation data
diverse. We also included images of birds from different videos available online with different resolutions.

348
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed

Figure 13: Frames from the Train and Validation set of Birds
4. Experimental Work
4.1. Final Model Training
Our final model, as mentioned in chapter 4, performed as well as the single-class classifier. We ran
60 epochs on the multi-class dataset. The dataset didn’t contain a background-subtracted image as the
model first removed the background and passed it to ConV2D for parallel detection. It was also trained
on the instances in which the background was superimposing the drone to test the modification of
background subtraction. Also, to distinguish between smaller drones and birds. The data imparity
mentioned earlier was evident here, looking at the volatile precision graph. But the loss functions have
decreased monotonically, whether it be the overall loss function or general class loss function. This
proves that our addition of the background-subtracted model to the YOLOv5 has trained well, and the
two models worked cordially. Providing more data with less class imbalance and having an equal
number of drones and birds may help solve this volatile behavior of precision. Our absolute precision
and recall scores were 0.954 and 0.973 on the training dataset.

Figure 14: Enter Caption

4.2. Raspberry Pi Deployment and Real-Time Detection

We also created a deployment file that we can run on Raspberry Pi. The file can use a camera
provided by the device (Pi camera) to identify drones. We have used an OpenCV based library to run
the camera of the device. To use the model, it directly imports it from GitHub using Pytorch’s hub class.
Then it returns the label and parameters of detection after passing a frame through the model. We have
also written a function that uses these labels to generate the camera’s bounding box and plot these boxes
on the screen using the camera. We can run our model in real-time to identify drones using Raspberry

349
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
and a monitor.

5. Performance Evaluation
The altered version of YOLOv5 was
able to produce decent results on the test
dataset also. With a recision of 78.2% our
implementation.
Figure 15: Result on the Test
Batch using the Altered Final
Model

Some objects whose spatial features are

similar to drones are being misidentified as
drones. This is because our dataset contains
less number of negative instances. Also, our
model was always able to identify white-
colored drones. It sometimes fails to
identify black-colored drones. This is yet
again due to most drones in train videos
being of white color. These are clear cases
of overfitting on training data.We also tried
to prune our model to improve its
performance. We used Pytorch library’s
utils class to prune the model. After pruning, we had the updated weights for our model. As we
mentioned earlier, the reason to do so was to avoid overfitting. We found a minor improvement in the
performance, and precision came to around 80

6. References
1. Sarvepalli, Sarat Kumar. (2015). Deep Learning in Neural Networks: The science behind an
Artificial Brain. 10.13140/RG.2.2.22512.71682.
2. F. Chollet, Deep Learning with Python, 1st. USA: Manning Publications Co., 2017, isbn: 1617294438.
3. Google Trends: https://trends.google.com/trends/explore?date=allq=dronedetection April 2020.
4. Rahaman, Muhammad Mahin, Md Ali, Md Hasanuzzaman, (2019). BHCDR: Real-Time Bangla
Handwritten Characters and Digits Recognition using Adopted Convolutional Neural Network.
10.13140/RG.2.2.28972.10881.
5. Weiming Hu, Tieniu Tan, Liang Wang, Steve Maybank, A survey on visual surveillance of object
motion and behaviors, IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications
andReviews), 34 (3) (2004) 334-352.
6. Github Repository of YOLOv5: https://github.com/ultralytics/yolov5
7. “Real-Time Drone Detection Using Deep Learning Approach”: Third International Conference,
8. MLICOM 2018, Hangzhou, China, July 6-8, 2018, Proceedings
9. “Establishing Drone Detection System by Using Deep Learning and YOLOv3 Formatting”, Pro-
ceedings of the 11th Annual International Conference on Industrial Engineering and Operations
Man- agement Singapore, March 7-11, 202.
10. “Automated Drone Detection Using YOLOv4” by Subroto Singha and Burchan Aydin

350
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com
SHODH SAGAR®
Darpan International Research Analysis
ISSN: 2321-3094 | Vol. 12 | Issue 3 | Jul-Sep 2024 | Peer Reviewed & Refereed
11. Wikipedia Page Pruning : https://en.wikipedia.org/wiki/Decision tree pruning
12. “Real-Time and Accurate Drone Detection in a Video with a Static Background” Ulzhalgas Sei-
daliyeva, Daryn Akhmetov, Lyazzat Ilipbayeva, Eric T. Matson Sensors (Basel) 2020 Jul; 20(14):
3856.Published online 2020 Jul 10.
13. S. Srigrarom and K. Hoe Chew, ”Hybrid motion-based object detection for detecting and track-
ing of small and fast moving drones,” 2020 International Conference on Unmanned Aircraft Systems
(ICUAS), 2020, pp. 615-621, doi: 10.1109/ICUAS48674.2020.9213912.
14. “Small Object Detection using Deep Learning”-Aleena Ajaz, Ayesha Salar, Tauseef Jamal, Asif
Ullah Khan
15. “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”
16. Alqaysi, Hiba et al. “A Temporal Boosted YOLO-Based Model for Birds Detection around
WindFarms.” Journal of imaging vol. 7,11 227. 27 Oct. 2021, doi:10.3390/jimaging7110227
17. Coluccia, A.; Fascista, A.; Schumann, A.; Sommer, L.; Dimou, A.; Zarpalas, D.; Méndez,
M.; de la Iglesia, D.; González, I.; Mercier, J.-P.; Gagné, G.; Mitra, A.; Rajashekar, S. Drone
vs. Bird Detection: Deep Learning Algorithms and Results from a Grand Challenge. Sensors 2021,
21, 2824.https://doi.org/10.3390/s21082824
18. Swamy, H. (2020). Unsupervised machine learning for feedback loop processing in cognitive
DevOps settings. Yingyong Jichu yu Gongcheng Kexue Xuebao/Journal of Basic Science and
Engineering, 17(1), 168-183. https://www.researchgate.net/publication/382654014
19. Swamy, H. (2024). Smart spending: Harnessing AI to optimize cloud cost management.
International Journal of Artificial Intelligence Research and Development (IJAIRD), 2(2), 40-55.
https://doi.org/10.5281/zenodo.13132258
20. Swamy, H. (2024). Leveraging AI for enhanced application service monitoring. International
Journal of Computer Engineering and Technology (IJCET), 15(4), 85-99.
https://doi.org/10.5281/zenodo.13131932
21. Unlu, E., Zenou, E., Riviere, N. et al. Deep learning-based strategies for the detection and track-
ing of drones using several cameras. IPSJ T Comput Vis Appl 11, 7 (2019).
https://doi.org/10.1186/s41074-019-0059-x

351
© 2024 Published by Shodh Sagar. This is a Gold Open Access article distributed under the terms of the Creative Commons License
[CC BY NC 4.0] and is available on https://dira.shodhsagar.com