低空无人机实时探测的GCB-YOLOv5s算法  

GCB-YOLOv5s algorithm for real-time detection of a low altitude UAV

作　　者：杨辉跃 容易圣 简钰洪 贾轶钧 YANG Huiyue;RONG Yisheng;JIAN Yuhong;JIA Yijun(Army Logistics Academy of PLA,Chongqing 401311,China)

机构地区：[1]陆军勤务学院,重庆401331

出　　处：《兵器装备工程学报》2023年第7期1-8,共8页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(61871402);重庆市教委科研项目(KJQN202012903)。

摘　　要：为应对无人机“黑飞”“滥飞”带来的安全威胁,迫切需要对无人机进行有效的实时探测与识别。然而由于低空无人机的灵活性、微小型化及多干扰因素等原因,传统的雷达探测等方法表现不佳。为此,基于机器视觉探测,提出了一种低空无人机实时探测的GCB-YOLOv5s算法。该算法针对经典YOLOv5算法运算速度难以满足高清实时处理的问题,使用轻量级GhostNet网络取代了YOLOv5骨干网络中的卷积运算,简化网络结构,大幅提高了计算速度;并且通过引入CA注意力机制,以及使用BiFPN双向加权特征金字塔替换颈部的PANet结构,在网络结构简化的基础上提升检测准确性。通过现地拍摄无人机在建筑、云层、树木、阴暗等不同复杂背景下的飞行姿态,结合公开数据集,对算法进行训练和测试实验。实验结果表明,GCB-YOLOv5s在参数量和浮点数计算量上均减少了近40%,且可达96.7%的精确率、96.4%的召回率和97.5%的平均精度。In order to deal with the security threat caused by “black flying” and “indiscriminate flying” of UAVs,it is urgent to carry out effective real-time detection and identification of UAVs.However,due to the flexibility,miniaturization and multiple interference factors of low altitude UAVs,the traditional radar detection methods are difficult to deal with.Therefore,based on machine vision detection,this paper proposes a GCB-YOLOv5s algorithm for low altitude UAV real-time detection.Aiming at the problem that the operation speed of the classical YOLOv5s algorithm is difficult to meet the high-definition real-time processing,the proposed algorithm uses a lightweight GhostNet network to replace the convolution operation in the YOLOv5s backbone network,which simplifies the network structure and greatly improves the calculation speed.The CA attention mechanism is also introduced in this algorithm.What's more,the PANet structure of the neck is replaced by the BiFPN bidirectional weighted feature pyramid.Then,the detection accuracy is improved on the basis of the simplification of the network structure.Subsequently,the flight attitudes of UAVs under different complex backgrounds such as buildings,clouds,trees and shadows are photographed.Combining public data sets,the algorithm is trained and tested.The test results show that the GCB-YOLOv5s algorithm reduces the number of parameters and the floating-point calculation amount by nearly 40%,and achieves an accuracy of 96.7%,a recall rate of 96.4% and an average accuracy of 97.5%.

关 键 词：低空无人机 目标识别 YOLOv5 实时探测 轻量化设计 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]