针对弱小无人机目标的轻量级目标检测算法  

Lightweight Target Detection Algorithm for Small and Weak Drone Targets

作　　者：蒋镕圻 叶泽聪 彭月平[2] 谢郭蓉 杜衡 Jiang Rongqi;Ye Zecong;Peng Yueping;Xie Guorong;Du Heng(Graduate Team,Engineering University of PAP,Xi’an,Shaanxi 710086,China;School of Information Engineering,Engineering University of PAP,Xi’an,Shaanxi 710086,China;School of Civil Engineering,Xinjiang University,Urumqi,Xinjiang 830000,China)

机构地区：[1]武警工程大学研究生大队,陕西西安710086 [2]武警工程大学信息工程学院,陕西西安710086 [3]新疆大学建筑工程学院,乌鲁木齐新疆830000

出　　处：《激光与光电子学进展》2022年第8期99-110,共12页Laser & Optoelectronics Progress

基　　金：武警工程大学科研创新团队课题(KYTD201803);武警工程大学基础研究项目(WJY201905)。

摘　　要：为解决无人机“滥用”带来的安全隐患,针对现有基于深度学习的无人机目标检测算法复杂度较高,导致模型训练耗时长、占用计算资源大、输入图像尺寸受限、检测速度慢等问题,提出了一种轻量级无人机目标检测(DTD-YOLOv4-tiny)算法。所提算法以YOLOv4-tiny为基础,通过K-means++聚类算法对Anchor box进行优化,并增加52×52尺寸特征图的检测头,拓展了算法对小目标的适用范围,再结合ShuffleNetv2轻量化骨干网络,使用reorg_layer下采样和sub-pixel上采样的方式,分别对YOLOv4-tiny算法的Backbone、Neck和Head进行优化,最终得到的模型大小仅为1.4 MB,浮点运算量(GFLOPs)仅为1.1的DTD-YOLOv4-tiny轻量级检测算法。实验结果表明,DTD-YOLOv4-tiny检测模型在不限制图像输入尺寸的同时,保证了较低的运算资源占用和高的检测实时性,同时降低参数量后的算法在面对原始大尺寸图像时也可以保持准确性。在Drone-vs-Bird 2017数据集上使用960×540尺寸的图像作为输入时,所提算法的平均精度(AP)@50值达到95%,在RTX2060显卡上的检测速度达到113 frame/s;在TIB-Net数据集上使用1920×1080尺寸的图像作为输入时,所提算法的AP@50值达到85.1%,在RTX2080Ti显卡上的检测速度达到119 frame/s。To address the security risks associated with drone“abuse”,aiming at the high complexity of the existing deep learningbased drone target detection algorithm,which results in lengthy model trainings,large computing resources,limited input image size,and slow detection speed,a lightweight level drone target detection(DTDYOLOv4-tiny)algorithm is proposed.The proposed algorithm is based on YOLOv4-tiny,and we optimized the Anchor box using the Kmeans++clustering algorithm,added the detection head of the 52×52 size feature map to expand the scope of the algorithm for small targets,and combined it with the ShuffleNetv2 lightweight backbone network,and the reorg_layer downsample and subpixel upsample methods were used to optimize the Backbone,Neck,and Head of the YOLOv4-tiny algorithm.Eventually,we obtained the DTDYOLOv4-tiny with a model size of 1.4 MB and a floatingpoint calculation(GFLOPs)of 1.1,which is a lightweight detection technique.The experiments demonstrate that the DTDYOLOv4-tiny detection model does not limit the image input size,while ensuring low computational resource occupation and high realtime detection.Simultaneously,the algorithm with reduced parameters can also maintain accuracy when facing the original largescale image.When using 960×540 size image as input on the DronevsBird 2017 dataset,the average precision(AP)@50 of the proposed algorithm achieved 95%,and the detection speed on the RTX2060 graphics card attained 113 frame/s;when using 1920×1080 size image as input on the TIBNet dataset,the AP@50 of the proposed algorithm achieved 85.1%,and the detection speed on the RTX2080Ti graphics card attained 119 frame/s.

关 键 词：图像处理 弱小无人机目标 DTD-YOLOv4-tiny 轻量级检测模型 实时目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]