基于改进YOLO-X模型及超分辨率增强的航拍机动车检测方法  

UAV VEHICLE DETECTION BASED ON IMPROVED YOLO-X MODEL AND SUPER-RESOLUTION ENHANCEMENT

作　　者：胡淳凯 陈泳 Hu Chunkai;Chen Yong(School of Aeronautics and Astronautics,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学航空航天学院,上海200240

出　　处：《计算机应用与软件》2025年第2期165-170,180,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(51875346)。

摘　　要：在无人机航拍数据集图像上提出基于改进YOLO-X的机动车识别算法,为工业巡检和交通监控业界提供思路。首次提出适用于局部特征样本的离线加在线分阶段数据增强方法。进一步改进焦点损失FocalLoss的同时,替换此前业界通用的交叉熵损失。重新设计目标检测Head组合,使模型检测精度mAP比业界通用模型提升0.042。建立在扩展感知域动态实时调整识别范围的算法,将运行时帧率提升19%。为航拍机动车算法实现纯ARM(Advanced RISC Machines)部署提供了可行性。We proposed a vehicle detection model based on improved YOLO-X on VisDrone2019 dataset for industrial inspections and traffic monitoring.This paper proposes an offline and online staged data augmentation method for local feature samples for the first time.The FocalLoss was further improved and replaced by the cross-entropy loss that was previously common in the industry.The target detection Head combination was redesigned,and the model detection accuracy mAP was 0.042 higher than the industry general model.The algorithm based on extended sensing domain dynamically adjusted the recognition range in real time,and the runtime frame rate was increased by 19%,which provided the feasibility for the implementation of ARM deployment of real-time vehicle detection.

关 键 词：深度学习 FocalLoss 目标检测 图像处理 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]