Drone vs. Bird Detection: Deep Learning Algorithms and Results from a Grand Challenge.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automatic detection of small drones is crucial for public and private sector security.
• Distinguishing drones from birds and background clutter presents a significant challenge in video analysis.

Purpose of the Study:

• To present and compare three novel deep learning approaches for the "Drone vs. Bird" detection problem.
• To evaluate algorithm performance in accurately identifying drones while minimizing false alarms from birds or other scene elements.

Main Methods:

• Development of three distinct deep learning strategies for drone detection.
• Comparison of algorithms on a real-world dataset from the 2020 Drone vs. Bird Detection Challenge.
• Evaluation metrics included correct detection rate, false alarm rate, and average precision.

Main Results:

• Performance varied based on drone visibility (size, shape) and environmental factors like camera movement.
• Sequences with moving cameras and distant drones proved most challenging.
• The three proposed deep learning approaches demonstrated complementary strengths in detection accuracy and false alarm reduction.

Conclusions:

• Deep learning offers promising solutions for automated drone detection.
• Challenges remain in robust detection under difficult conditions such as camera motion and long-range observation.
• Complementary algorithm performance suggests potential for ensemble methods in enhancing drone detection systems.