AI-Based Drone Detection for Border Security Using Deep Learning Aris | Rafizi

Background

With the increasing popularity of drone usage across various sectors, such as delivery services and surveillance, there is also a growing concern regarding the potential misuse of drones for illegal activities. In sensitive areas, such as national borders, unauthorized drone activity poses significant security risks. These drones can be used for smuggling, surveillance, or other criminal purposes, requiring authorities to develop advanced detection systems to mitigate these threats.

Problem

Unauthorized drone intrusions along national borders have become a growing issue, as drones can easily evade traditional surveillance methods. This poses a significant security threat, especially for military and law enforcement agencies tasked with protecting sensitive areas. A robust system is needed to accurately detect and track drones in real-time to prevent illegal activities.

Objective

The objective of this project is to develop a real-time drone detection system using deep learning techniques that can be deployed in border surveillance systems, such as CCTV. The system will be based on object classification models like CNN and ResNet50, capable of identifying drones in live video feeds and alerting security personnel to take immediate action. This solution will help enhance border security by quickly detecting and neutralizing potential drone threats.

Datasets

The dataset was collected from Aerodyne Company and comprises approximately 1,625 images, including drone and non-drone pictures. The drone images feature two types of drones: multi-rotors and fixed-wing drones. Non-drone images include birds, airplanes, and other objects. The primary objective of this model is to classify objects near the border as either drones or non-drones.

Figure 1: Sample of datasets.

Data Analysis

The dataset was analyzed to derive insights, with data visualization techniques such as pie charts used to illustrate the percentage distribution of subjects labeled as drones and non-drones. An example of this analysis is presented in Figure 2, showcasing the proportions in a clear and intuitive format.

Figure 2: Class distribution in Datasets.
Figure 3: Sample of visualize images from the training Dataset.
Figure 4: Features of Datasets.

Data Modelling

Two predictive models were developed using Convolutional Neural Networks (CNN) and ResNet50 architectures. The dataset was partitioned into training and testing sets, with 70% allocated for training to teach the models and 30% reserved for testing to evaluate their performance. These models aim to accurately classify the dataset into drone and non-drone categories, leveraging the strengths of CNN for feature extraction and ResNet50 for handling deep network structures efficiently.

Table 1: Parameters for the models.

The results for the models trained as below:

Figure 5: Result of classification by using CNN.

Based on Figure 5, we can conclude that the model correctly classifies 85% of the total samples (494 samples). It performs exceptionally well in predicting the Drone class, achieving a high precision of 0.95 and a balanced F1-score of 0.89. However, the model is less effective for the Non-Drone class, showing lower precision at 0.70.

Figure 6: Result of classification by using ResNet50

The model demonstrates excellent performance overall, achieving a high accuracy of 94%. Both classes exhibit strong precision, recall, and F1-scores, indicating the model is effective at distinguishing between Drone and Non-Drone classes.

Final Conclusion:

ResNet50 is the superior model, achieving higher accuracy (94%), better precision, recall, and F1-scores for both Drone and Non-Drone classes.

ResNet50's performance improvement is particularly significant in the Non-Drone class, where CNN struggled with lower precision.

While CNN provides acceptable results, ResNet50 is more reliable and effective for this classification task, making it the recommended model for deployment.