GA-Fly

The proliferation of drones across various sectors has increased the need for robust counter-drone systems to mitigate security and privacy risks. This study introduces YOLO Drone, an edge-deployable, efficient drone tracking and aiming algorithm based on YOLOv8, which is designed to enhance real time ground-to-air drone detection. The algorithm addresses the challenges of detecting small, fast-moving drones by incorporating multiscale dense connection (MDC) and multilayer shared head (MSH), optimizing YOLOv8 for superior accuracy in small target detection. To ensure operational viability on edge platforms with limited computational resources, we implement model pruning and distillation, resulting in a compact model that maintains high performance with a reduced computational load. The system also integrates a gimbal camera tracking and aiming system for dynamic target acquisition, significantly expanding the detection range beyond fixed camera limitations. The experiments revealed that YOLO-Drone improved the detection accuracy; the accuracy was 94.3%, the recall rate was 88.2%, the mAP50 was 93%, and the MAP50:95 was 54.9%, which were 0.1%, 19%, 9.6% and 6.7% higher than those of the baseline YOLOv8 model. The optimized model, deployed on the NVIDIA Jetson AGX Orin, achieves real-time processing, showcasing its potential for practical edge deployment. This research contributes a comprehensive framework for drone detection, offering a feasible solution for real-time, accurate drone tracking and aiming at edge devices. The dataset and demo video are available at https://github.com/ballballubsmart/YOLO-Drone 

随着无人机在各行业的普及，相关方对高性能反无人机系统的需求日益攀升，以应对安全与隐私风险。本研究提出了基于YOLOv8的边缘可部署高效无人机跟踪瞄准算法YOLO Drone，旨在提升地面对空无人机的实时检测能力。该算法通过融入多尺度密集连接（Multiscale Dense Connection, MDC）与多层共享头（Multilayer Shared Head, MSH）对YOLOv8进行优化，从而提升小目标检测精度，解决小型高速移动无人机的检测挑战。为确保该算法在计算资源有限的边缘平台上稳定运行，研究团队采用模型剪枝与知识蒸馏技术，得到一款轻量化模型，在降低计算负载的同时仍保持高性能。该系统还集成了云台相机跟踪瞄准模块，用于动态目标捕获，突破固定相机的检测范围限制，大幅拓展了检测视野。实验结果显示，YOLO-Drone的检测性能得到显著优化：其精确率达94.3%，召回率为88.2%，mAP50为93%，mAP50:95为54.9%，较基准YOLOv8模型分别提升0.1%、19%、9.6%与6.7%。经优化后的模型部署于英伟达（NVIDIA）Jetson AGX Orin平台时可实现实时处理，展现了其在实际边缘场景中部署的潜力。本研究构建了一套完整的无人机检测框架，为边缘设备上的实时精准无人机跟踪瞄准提供了可行解决方案。本研究的数据集与演示视频可通过以下链接获取：https://github.com/ballballubsmart/YOLO-Drone

搜集汇总

背景与挑战

背景概述

GA-Fly是一个用于无人机小目标检测的数据集，包含10,800张高清图像及其标注文件，涵盖多种拍摄角度和光照条件。数据集目前仅公开测试集，其余数据将在文章发表后补充。