Unauthorized Drone Detection Using Deep Learning

P Bhavani, Prateek Prasanna, M Dhivagar, S VelliangiryInternational Journal for Research in Applied Science and Engineering Technology 2024 · DOI

TL;DR

This paper presents a real-time drone detection system leveraging the YOLOv8 deep learning architecture to combat security risks like unauthorized surveillance and smuggling. By fine-tuning the model on a large, annotated dataset of nearly 10,000 images, the researchers developed a solution capable of identifying small, agile UAVs against complex backgrounds with high precision and speed.

What Problem Does This Paper Address?

As Unmanned Aerial Vehicles (UAVs)—commonly known as drones—become more accessible, they pose significant security and privacy challenges. From illegal drug smuggling across borders to unauthorized espionage in restricted airspace, the misuse of drones is a growing concern.

Traditional detection methods often fall short. Radar systems struggle to differentiate small, low-flying drones from birds or other clutter, while radio frequency (RF) and acoustic sensors can be hampered by environmental noise. Computer vision offers a promising alternative, but it faces its own hurdles: drones are small, move quickly, and must be detected in real-time against "noisy" visual backgrounds like trees, buildings, or clouds. The authors aim to bridge this gap by implementing a state-of-the-art object detection algorithm tailored for the unique visual signatures of drones.

Methods

The researchers utilized YOLOv8 (You Only Look Once version 8), a cutting-edge single-pass object detection model. Unlike older algorithms that scan an image multiple times, YOLOv8 processes the entire image in one go, making it exceptionally fast and suitable for real-time applications.

Key steps in their methodology included:

  • Transfer Learning: Leveraging pre-trained weights to allow the model to adapt quickly to the specific features of drones.
  • Data Augmentation: To prevent overfitting and improve the model's ability to generalize, the authors applied techniques like rotation, flipping, and brightness adjustment to the training data.
  • Multi-scale Detection: YOLOv8’s architecture allows it to detect objects of varying sizes, which is critical for identifying drones at different distances from the camera.
  • Pre-processing: Images were cleaned, resized, and normalized to ensure consistent input for the neural network.

The abstract also mentions the potential integration of Kalman filtering for post-processing to refine tracking and reduce false positives, though the core of the paper focuses on the detection performance of the YOLOv8 model itself.

Datasets

The study utilized an open-source dataset sourced from Kaggle, specifically the "YOLOv8 - YOLO Drone Detection Dataset" created by user marquis03.

  • Size: 9,939 images and corresponding labels.
  • Content: The dataset includes annotated images and videos of drones in diverse environments and aerial perspectives, providing the variety necessary for robust training.

Key Insights and Ideas

  • The Speed-Accuracy Trade-off: The choice of YOLOv8 highlights the necessity of "single-pass" detection. In security contexts, a delay of even a few seconds can mean the difference between a successful interception and a security breach.
  • Contextual Robustness: One of the "aha" moments in the research is the model's ability to maintain high precision despite background noise. By training on diverse environments, the model learns to ignore "distractor" objects (like birds or planes) that often trip up simpler systems.
  • The Importance of Confidence Thresholds: The researchers provide detailed Precision-Recall and F1-score curves. These curves illustrate that the model's peak performance (F1 score) occurs at a specific confidence threshold, teaching us that "tuning" the model's sensitivity is just as important as the training itself.

Key Results

The paper evaluates the model using standard computer vision metrics:

  • F1 Score: The model achieved a peak F1 score of 0.84 at a confidence threshold of 0.383. This represents a strong balance between precision (avoiding false alarms) and recall (not missing any drones).
  • Precision and Recall: The Precision-Confidence curve shows that as the confidence threshold increases, precision approaches 1.0, indicating the model becomes highly reliable in its "certain" detections.
  • Real-time Capability: The authors emphasize that the system is capable of real-time prediction, which is vital for immediate response to unauthorized intrusions.

Limitations and Future Work

While the results are promising, the authors acknowledge several areas for improvement:

  • Dataset Diversity: The current model is trained primarily on visual data. The authors suggest expanding the dataset to include a wider variety of drone types and even integrating non-drone sounds to create a multi-modal detection system.
  • Real-world Deployment: The study was conducted on a curated dataset. Future work involves deploying the system in unpredictable, real-world environments to test its reliability under varying weather and lighting conditions.
  • Hardware Optimization: The authors aim to further reduce hardware requirements to make the system more cost-effective and scalable for widespread surveillance use.

Summary updated

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.
BibTeX
@article{bhavani2024unauthorized,
  author = {P Bhavani and Prateek Prasanna and M Dhivagar and S Velliangiry},
  title = {{Unauthorized Drone Detection Using Deep Learning}},
  journal = {International Journal for Research in Applied Science and Engineering Technology},
  year = {2024},
  doi = {10.22214/ijraset.2024.61496},
}
Figure