Drone Model Classification Using Convolutional Neural Network Trained on Synthetic Data.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Automated drone identification is crucial for various applications, including security and traffic management.
• Manual labeling of drone footage is time-consuming and labor-intensive.
• Existing methods may struggle with variations in lighting, angle, and background in real-world scenarios.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for accurate drone model identification in real-life videos.
• To demonstrate the effectiveness of using synthetic data with domain randomization for training drone detection models.
• To reduce the time and cost associated with manual drone data labeling.

Main Methods:

• Generated synthetic drone images using domain randomization to vary textures, backgrounds, and orientations.
• Trained a CNN model on the synthetic dataset to classify three common drone models: DJI Phantom, DJI Mavic, and DJI Inspire.
• Validated the model's performance on a real-life drone video dataset (Anti-UAV).

Main Results:

• The CNN achieved high performance on the real-life dataset, with an overall accuracy of 92.4%.
• Key performance metrics included a precision of 88.8%, recall of 88.6%, and an F1 score of 88.7%.
• The study confirmed the transferability of models trained on synthetic data to real-world video analysis.

Conclusions:

• The proposed CNN model effectively identifies specific drone models in real-life videos.
• Synthetic data generation with domain randomization is a viable and efficient approach for training robust drone identification systems.
• This method offers a significant reduction in manual labeling effort and proves effective for real-world drone surveillance applications.