An Integrated AI-Powered Air Defense System for Threat Detection and Neutralization

Abstract

The growing complexity of low-altitude airspace and the rapid proliferation of unmanned aerial vehicles have created an urgent demand for automated systems capable of identifying aerial threats with speed and precision. This project presents the design and development of an integrated AI-powered aerial object detection and classification system, trained on two datasets spanning seven distinct classes: drones, helicopters, airplanes, birds, military airplanes, missiles, and UAVs. Three deep learning architectures were evaluated for image classification — YOLOv26x, EffcientNetV2-L, and Swin Transformer — achieving accuracies of 99.40%, 98.72%, and 99.19% respectively on the custom DAOD dataset. For spatial object detection, YOLOv26x combined with Slicing Aided Hyper Inference (SAHI) achieved an mAP@0.5 of 0.981 on the AOD4 dataset and 0.931 on the more complex DAOD dataset, demonstrating strong generalization across varying target sizes and backgrounds. The system addresses key challenges including class imbalance, inter-class visual similarity between drones and birds, and small object detection at long range. Results confirm that a well-regularized, medium-scale YOLO architecture provides the optimal balance between detection accuracy and computational effciency for realtime air defense applications