Aerial Supervision of Drones and Other Flying Objects Using Convolutional Neural Networks

Accurate detection of distant drones in clustered environment amidst other flying objects such as birds is of critical importance in anti-drone system design. This study proposed a novel object detection model that efficiently detect and differentiate drones from other flying objects under different weather conditions. The custom dataset consists of manually generated drone images and bird samples under sunny, cloudy and evening conditions. The simulation result shows that KITYOLO outperformed YOLOv5 both in precision (sunny 96.2% vs 85%; cloudy 73.7% vs 26.3%; evening 58.5% vs 26.1%) and recall (evening 42.4% vs 15%) in all aspects with an overall F1-score of 98% as against 91.9% while maintaining timeliness and memory usage.