12 V May 2024

https://doi.org/10.22214/ijraset.2024.61496
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

Unauthorized Drone Detection Using Deep

Learning
Bhavani P1, Prasanna P2, Dhivagar M3, Velliangiry S4
Sri Manakula Vinayagar Engineering College

Abstract: The growing concerns over drone misuse in national airspace, including activities like drug smuggling and privacy
violations, demand advanced drone detection systems. Traditional methods struggle to accurately identify drones among various
airborne objects. YOLO v8, renowned for its efficient single-pass object detection, presents a sophisticated solution. Through
training on annotated datasets featuring drones in diverse environments and fine-tuning with aerial imagery, YOLO v8 excels at
recognizing drones amidst complex backgrounds. Its multi-scale detection capabilities enable it to detect drones of varying sizes,
crucial for different distances from the camera. Real-time prediction ensures timely detection, vital for immediate responses to
unauthorized intrusions. Integration with drone tracking systems enhances situational awareness, facilitating proactive
measures. Furthermore, employing adaptive thresholding and post-processing techniques like Kalman filtering refines drone
detection accuracy, reducing false positives. YOLO v8 thus emerges as a potent tool in mitigating risks associated with drone
misuse, bolstering defenses against unauthorized activities in national airspace.
Keywords: YOLO v8, OpenCV, robust defense, datasets.

I. INTRODUCTION
Drones, known as unmanned aerial vehicles (UAVs), have gained popularity across various sectors for their diverse applications,
from military reconnaissance to commercial photography and security surveillance. However, their widespread use also poses
security risks, including espionage and privacy breaches. Detecting drones presents a complex challenge due to their small size and
agility. Traditional radar systems often struggle to differentiate drones from other objects. To address this, advanced technologies
like radio frequency (RF) sensors, acoustic detection, and computer vision algorithms are being employed. Integration of these
technologies enhances accuracy, with artificial intelligence, particularly deep learning, enabling adaptation to evolving drone
technologies. Thus, a multifaceted approach combining various detection methods offers a comprehensive solution to mitigate
security threats posed by drones in both military and commercial contexts.
Integrated drone detection systems, employing RF sensing, acoustic detection, and computer vision enhanced by AI, are vital for
accurate identification across diverse scenarios. Privacy concerns mandate anonymization of data to respect individual rights.
Regulatory frameworks, including operator registration and no-fly zones, are crucial for managing drone-related risks. Collaboration
between stakeholders is essential for effective regulation and technology development. Continuous research and development
efforts are needed to stay ahead of evolving threats, ensuring the dynamic evolution of counter-drone technologies. In summary,
while drones offer benefits, robust detection mechanisms, balanced security-privacy approaches, and effective regulations are
imperative to manage associated risks effectively.

II. RELATED WORK

1) Real-Time Detection and Tracking of Multiple Drones with Deep Neural Networks" by Y. Hu, M. Lu, and Y. Li [1]
The research paper introduces a novel real-time drone detection and tracking system utilizing deep neural networks. Integrating
YOLOv3 for detection and DeepSORT for tracking, the system is trained on a diverse dataset including authentic and synthetic
drone data. Testing on drone videos showcases high accuracy in detection and tracking. Unique to this system is the combination of
YOLOv3 and DeepSORT, leading to superior speed and accuracy compared to existing methods. The authors emphasize its
versatility, suggesting applications in drone surveillance and monitoring. By addressing detection and tracking challenges in real-
time, the system enhances security and situational awareness in areas requiring vigilant drone monitoring.

2) Drone Detection and Tracking Using Convolutional Neural Networks and Optical Flow by M. Hahn and K. Leppkes [2]
The paper presents a real-time drone detection and tracking system using a combination of CNNs and optical flow, optimized for
single-camera setups.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 282
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

The deep learning model is trained on a diverse dataset of drone images and videos, including synthetic and real-world data. A
novel tracking algorithm incorporating Kalman filtering and the Hungarian algorithm enhances precision. Evaluation on real-world
drone videos demonstrates high accuracy in detection and tracking. The synergy between CNNs and optical flow, along with the
innovative tracking algorithm, ensures effectiveness in real-time scenarios. This research sets the stage for robust drone monitoring
systems with single-camera setups, applicable in security, surveillance, and situational awareness contexts where timely and
accurate drone detection and tracking are crucial.

3) Detection and Classification of UAVs Using Deep Learning and Feature Extraction Techniques by A. Khawar and S. Shahid [3]
The paper presents a drone detection and classification system integrating deep learning with feature extraction techniques,
particularly employing Convolutional Neural Networks (CNNs) for classification and Principal Component Analysis (PCA) for data
reduction. Through CNNs, the model learns intricate patterns from drone images, enabling accurate classification. PCA enhances
system efficiency by reducing input data complexity. Testing on a dataset of drone images shows high accuracy in both detection
and classification, demonstrating the system's effectiveness in discerning and categorizing drones based on visual features. The
system's versatility suggests potential applications in drone surveillance and security. Its accurate detection and classification
capabilities address concerns regarding unauthorized drone activities, offering valuable support for security measures in various
environments. Overall, the paper provides a robust solution merging deep learning and feature extraction for drone detection and
classification.

4) Automated drone detection and tracking system using deep learning" by S. Farhan, M. Waqar, and F. Ahmed [4]
The research introduces an automated drone detection and tracking system, combining YOLOv3 for detection and deep association
network (DAN) for tracking. Trained on real-world data, the system exhibits high accuracy in both tasks, showcasing its
effectiveness in practical scenarios. The integrated approach offers potential applications in border security and surveillance,
providing timely and accurate detection and tracking of drones. Empowered by deep learning algorithms, the system contributes to
enhancing security measures where drone detection and tracking are crucial, emphasizing real-world applicability and accuracy.

5) Detection and tracking of low-altitude UAVs using deep learning- based techniques" by T. Kim, Y. Kim, and J. Kim in
Aerospace Science and Technology [5]
The research paper introduces a novel system for detecting and tracking low-altitude Unmanned Aerial Vehicles (UAVs) using
deep learning techniques. It combines Convolutional Neural Networks (CNNs) for spatial feature extraction and Recurrent Neural
Networks (RNNs) for capturing temporal dependencies in video frames, enabling dynamic analysis of UAV movements. A unique
aspect is the introduction of an algorithm to predict UAV flight paths, enhancing proactive tracking. Evaluation on real-world UAV
videos demonstrates high accuracy in detection and tracking tasks. The emphasis on real-world data enhances the system's practical
applicability, suggesting its potential in critical scenarios requiring precise UAV monitoring. Overall, the paper contributes to UAV
surveillance technologies by leveraging deep learning and predictive algorithms for enhanced detection and tracking capabilities.

III. IMPLEMENTATION
A. Data Collection
The data utilized in this research is sourced from the Kaggle open-source platform, specifically from a dataset titled "YOLOv8 -
YOLO Drone Detection Dataset" created by a user named marquis03. This dataset likely contains images and videos relevant to
drone detection, curated specifically for training and testing purposes. Given the name "YOLOv8," it suggests a focus on utilizing
the YOLO (You Only Look Once) algorithm, a popular choice for object detection tasks, including drone detection. Such datasets
typically include annotated images or videos where drones are labeled or marked, allowing algorithms to learn and identify drones
accurately. By leveraging open-source datasets like this from platforms such as Kaggle, researchers and developers can access
valuable resources to train and evaluate their drone detection models, contributing to advancements in the field of aerial surveillance
and security.
https://www.kaggle.com/code/marquis03/yolov8-yolo-drone-detection-dataset/data
Pre-processing and splitting the data set:
Pre-processing and splitting are crucial steps in preparing data for machine learning tasks, including drone detection. Pre-processing
involves cleaning and transforming raw data to a suitable format for analysis. In the context of drone detection datasets, this may
include tasks such as resizing images, removing noise, or normalizing pixel values to ensure uniformity and consistency.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 283
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

Additionally, pre-processing may involve augmenting the dataset by applying techniques like rotation, flipping, or adjusting
brightness to increase its diversity and improve model generalization. After pre-processing, the dataset is typically split into training,
validation, and testing sets.

The training set is used to train the model, the validation set is employed to fine-tune model hyperparameters and monitor
performance during training, while the testing set assesses the final model's performance on unseen data. The splitting process
ensures that the model is trained on a diverse range of data and evaluated rigorously, helping to prevent overfitting and providing
reliable estimates of its generalization capabilities. Overall, pre-processing and splitting are essential steps in the data pipeline for
drone detection, facilitating effective model training and evaluation.

B. Model Creation
Model creation using the YOLO (You Only Look Once) algorithm involves several key steps to build an effective object detection
system, including for drone detection. Firstly, the dataset comprising images or videos containing drones and annotations indicating
their locations must be prepared. Next, the YOLO architecture, which typically consists of convolutional layers followed by fully
connected layers, is configured. These layers extract features from the input images and predict bounding boxes and class
probabilities for detected objects. Overall, model creation using the YOLO algorithm is a comprehensive process involving data
preparation, architecture configuration, training, and deployment, ultimately leading to robust and efficient drone detection systems.

IV. RESULT AND DISCUSSION

In the context of object detection using the YOLOv8 algorithm, the results section typically includes the performance metrics such
as accuracy, precision, recall, and F1 score, obtained from evaluating the trained model on test datasets. Additionally, it may include
qualitative assessments, such as visualizations of detected objects overlaid on input images or videos. The discussion section then
interprets and contextualizes the results presented in the preceding section. Here, researchers analyze the performance metrics,
comparing them with those of other models or approaches. They discuss the strengths and limitations of the YOLOv8 algorithm,
considering factors such as computational efficiency, accuracy, robustness, and scalability. Any observed trends or patterns in the
results are explored, along with potential reasons for them. Furthermore, the discussion section may address the implications of the
results for real-world applications, potential areas for further research or improvement, and any practical considerations or
challenges encountered during the experimentation process. It provides insights into the significance of the findings, their relevance
to the broader field of object detection, and their potential impact on future research or applications. Overall, the results and
discussion sections work together to communicate the outcomes of the study, provide context and interpretation of the results, and
offer insights into their implications and significance. They are crucial for conveying the research findings effectively and
contributing to the advancement of knowledge in the field.

A. Accuracy
In the context of YOLOv8, accuracy still measures the correctness of the model's predictions. However, the application of accuracy
can differ slightly due to the nature of object detection tasks. YOLOv8 is primarily used for object detection, where the model is
trained to detect and localize objects within an image or video. In this scenario, accuracy refers to the proportion of correctly
detected objects (both true positives and true negatives) relative to the total number of objects in the dataset. True positives
represent objects correctly detected by the model, while true negatives denote areas where no object is present, and the model
correctly identifies them as such.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 284
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

Therefore, accuracy in YOLOv8 reflects the model's ability to accurately detect objects within the given dataset, providing an
overall measure of the model's performance in terms of object detection accuracy.

B. Precision
In YOLOv8, precision assesses the accuracy of positive predictions, particularly regarding the detection of objects. Precision
calculates the proportion of correctly identified objects (true positives) among all the instances predicted as positive by the model. In
object detection tasks, true positives represent instances where the model correctly identifies and localizes objects within the image
or video frame.

Precision focuses solely on the correctness of positive predictions, emphasizing the model's ability to avoid falsely identifying
background regions as objects. A high precision score indicates that the model makes fewer false positive predictions, providing a
more reliable identification of objects within the dataset. Therefore, precision in YOLOv8 reflects the model's ability to accurately
pinpoint objects of interest amidst the background noise, contributing to its overall performance in object detection tasks.

C. Recall
In YOLOv8, recall, also referred to as sensitivity or true positive rate, evaluates the model's capability to correctly identify all
positive instances within a dataset. It measures the proportion of true positive predictions, which are instances correctly identified as
positive by the model, relative to the total number of actual positive instances present in the dataset. In the context of object
detection tasks, true positives represent objects correctly detected and localized by the model. Therefore, recall focuses on the
model's effectiveness in capturing and recognizing all instances of the target objects within the dataset, regardless of whether there
are any false positives or false negatives.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 285
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

A higher recall score indicates that the model can effectively identify a larger portion of the positive instances, suggesting better
coverage and comprehensiveness in detecting objects of interest. Thus, recall in YOLOv8 provides insights into the model's
performance in terms of accurately identifying relevant objects within the input data.

D. F1 SCORE
In YOLOv8, the F1 score serves as a balanced metric that takes into account both precision and recall simultaneously. It is
calculated as the harmonic mean of precision and recall, combining the strengths of both metrics. By incorporating both precision
and recall, the F1 score provides a comprehensive assessment of the model's performance in object detection tasks. A high F1 score
indicates that the model achieves both high precision (few false positives) and high recall (few false negatives), striking a balance
between correctly identifying objects and avoiding misclassifications.

This balanced measure is particularly useful when dealing with imbalanced datasets or when both precision and recall are equally
important for the task at hand. Therefore, in YOLOv8, the F1 score offers valuable insights into the overall effectiveness and
reliability of the model in detecting and localizing objects within images or video frames.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 286
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 12 Issue V May 2024- Available at www.ijraset.com

V. CONCLUSION AND FUTURE WORK

In conclusion, the proposed drone detection and classification system presents a significant advancement by integrating computer
vision techniques, deep learning algorithms, and the YOLOv8 algorithm. Its dual-module architecture enables comprehensive
analysis of drone objects in real-time video streams, bolstered by transfer learning for adaptability across diverse drone types and
environments. With reduced hardware requirements and faster processing times, the system is cost-effective and scalable, making it
suitable for time-sensitive applications like surveillance and security. Overall, this system offers a promising solution for enhancing
drone detection technologies, with potential to significantly improve capabilities in real-world scenarios requiring accurate
identification and classification of drones.
Increasing the size and diversity of the dataset: While the current system is trained on a dataset of drone and non-drone sounds,
expanding the dataset to include a wider variety of drones and non-drone sounds could improve the accuracy and robustness of the
model. Testing and validation in real-world environments: To truly evaluate the efficacy of the system, it will be necessary to
deploy it in real-world settings and assess its accuracy and reliability under different conditions.

REFERENCES
[1] "Real-Time Detection and Tracking of Multiple Drones with Deep Neural Networks" by Y. Hu, M. Lu, and Y. Li in IEEE Transactions on Industrial
Informatics.
[2] "Drone Detection and Tracking Using Convolutional Neural Networks and Optical Flow" by M. Hahn and K. Leppkes in IEEE International Conference on
Robotics andAutomation.
[3] "Detection and Classification of UAVs Using Deep Learning and Feature Extraction Techniques" by A. Khawar and S. Shahid in Sensors.
[4] "Automated drone detection and tracking system using deep learning" by S. Farhan,M. Waqar, and F. Ahmed in Journal of Ambient Intelligence and
Humanized Computing.
[5] "Detection and tracking of low-altitude UAVs using deep learning-based techniques"by T. Kim, Y. Kim, and J. Kim in Aerospace Science and Technology.
[6] "Real-time UAV detection using deep learning and YOLO algorithm" by S. Kim, S. Cho, and J. Kim in IEEE International Conference on Big Data and
Smart Computing.
[7] "Drone Detection and Tracking with Convolutional Neural Networks" by L. F. Rocha,F. L. D. Oliveira, and R. S. Ferreira in IEEE Latin American
Conference onComputational Intelligence.
[8] "Drone Detection Using Deep Learning Techniques" by V. L. Arévalo, L. Gómez, andJ. A. García in Journal of Sensors.
[9] "Deep learning-based UAV detection using synthetic and real-world data" by K. E. Lee, S. K. Hong, and H. J. Kim in Journal of Sensors
[10] "Detection and tracking of UAVs using deep learning and computer vision techniques" by S. S. Choudhary, S. M. Ali, and M. A. Abidi in IEEE International
Conference on Robotics and Biomimetics.

© IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 287