Real-Time UAV Detection Using an Enhanced YOLO v8 Model

Unmanned Aerial Vehicles (UAVs) have expanded beyond military applications due to advancements in the UAV industry. However, their widespread civilian use poses risks, as highlighted by the 2018 Gatwick Airport incident, where a drone disrupted flights, affecting 140,000 travelers. Accurate detection of small UAVs, which operate at low altitudes, slow speeds, and compact sizes, is crucial for urban safety. This study employs the You Only Look Once (YOLO) framework for precise UAV detection. Given the scarcity of publicly available UAV-specific datasets, UAV videos were manually collected under diverse conditions to develop a comprehensive dataset for model training. Three YOLO variants-YOLOv5, YOLOv8, and YOLOv10-were evaluated on the constructed UAV dataset. Among these, YOLOv8 demonstrated better performance. To further optimize detection accuracy, various learning rate (LR) scheduling techniques were explored. The implementation of a cyclical LR schedule yielded the best results, improving the mean average precision (mAP-50) and reducing the training time by 0.5 hours.