DeepVision: Enhanced Drone Detection and Recognition in Visible Imagery through Deep Learning Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Aerospace Engineering

Background:

• Drones pose significant risks to airport infrastructure due to misuse.
• Unauthorized drone activity has repeatedly disrupted flight operations.
• Distinguishing drones from birds is challenging due to similar appearances and flight patterns.

Purpose of the Study:

• To develop and evaluate an innovative deep learning method for distinguishing drones from birds.
• To address the challenge of detecting small drones and differentiating them from avian targets.
• To improve the accuracy and efficiency of drone detection systems for aviation security.

Main Methods:

• A novel deep learning approach was proposed for drone and bird classification.
• A robust image-tiling technique with overlaps was developed to enhance detection of small drones.
• Various convolutional neural network (CNN) models, including SqueezeNet, MobileNetV2, ResNet18, and ResNet50, were evaluated.
• Performance was assessed using a dataset from the 2020 Drone vs. Bird Detection Challenge and an unseen dataset.

Main Results:

• The SqueezeNet model demonstrated superior performance, achieving an average precision (AP) of 0.770 for medium area ratios.
• The proposed method, particularly with SqueezeNet, surpassed existing drone detection systems in accuracy and speed.
• The image-tiling technique significantly improved the detection of very small drones.
• The system accurately identified and differentiated between drones and birds.

Conclusions:

• The developed deep learning method offers a reliable solution for real-time drone detection and identification.
• SqueezeNet is highly suitable for accurate and efficient drone detection in airport environments.
• The approach enhances aviation security by mitigating risks associated with unauthorized drone presence.