Investigating Domain Adaptation Feasibility for Drone Detection: A CPU-based YOLO Approach using RGB-Infrared Images

Abstract

Drone technology has become crucial in security and national defense, with anti-drone systems playing a vital role in government and military operations. A key component of these systems is drone detection using infrared camera imagery. While deep learning represents the state-of-the-art approach for object detection, it requires extensive datasets for practical implementation. Given the limited availability of infrared image datasets, leveraging larger RGB image datasets through domain adaptation could potentially enhance detection capabilities. This study investigates the feasibility of RGB-Infrared domain adaptation for drone detection, implementing CPU-based processing across various YOLO models (YOLOv5n/x, YOLOv10n/x, and YOLOv11n/x). We trained twelve models using either RGB or infrared datasets and evaluated their performance both with and without domain adaptation. Without domain adaptation, the models achieved excellent mean average precision (mAP50) values exceeding 95% at speeds of 0.18 – 10.65 frames per second (FPS). With domain adaptation, RGB-trained models detecting drones in infrared images achieved mAP50 values of 42.6 – 52.4% at 0.18 – 9.08 FPS, while infrared-trained models failed to detect drones in RGB images. Our findings demonstrate that (1) YOLO models excel at drone detection given sufficient data, (2) features learned from RGB images can be adapted for infrared image detection but not vice versa, and (3) domain adaptation with CPU-based processing is feasible for drone detection applications.