UAVCity-YOLO: An Improved YOLOv13 Architecture for Robust UAV Detection Using Infrared Data in Noisy Urban Environments

Abstract

The rapid growth of unmanned aerial vehicles (UAVs), especially drones, in urban airspace raises significant security and safety concerns, creating an urgent demand for accurate real-time detection. While infrared imaging enables continuous surveillance, urban infrared UAV detection remains challenging due to the small target size, weak thermal contrast, and severe background thermal noise, which limit the robustness of existing detectors. This paper proposes UAVCity-YOLO, a lightweight and robust single-stage detector based on the YOLOv13n (You Only Look One version 13n) architecture, specifically designed for UAV detection using infrared data in thermally complex urban environments. The proposed approach integrates a coordinate-aware backbone for enhanced small-target representation, an improved multi-scale feature fusion Neck structure, and a novel infrared-aware regression loss to improve localization stability under thermal noise. Extensive experiments on urban infrared UAV datasets demonstrate that UAVCity-YOLO consistently outperforms state-of-the-art models. In particular, on the thermally noisy dataset, UAVCity-YOLO achieves an mAP0.5 of 89.1% and mAP0.5:0.95 of 57.9%, while maintaining fast inference at 1.5 ms and a compact model size of only 4.1 MB, confirming its effectiveness for real-time UAV surveillance.