摘要
高铁长大桥梁施工过程中,工地场景复杂多变,针对行业级无人机在高空巡检时,识别地面施工人员及安全帽等小目标精度差的问题,提出基于实时目标检测模型(RT-DETR)的优化智能检测方法,设计轻量级Mobile-HGNet来替代RT-DETR的主干网络,从而有效减少模型的计算量;提出级联分组自注意力模块,不仅能够保证模型的轻量化,还能增强全局和局部的区域关注,提升模型的检测性能。利用切片辅助超推理(SAHI)输入进行数据增强,以牺牲少量实时性换取更高的检测率。将改进后的RT-DETR模型与基线模型进行对比,优化后的RT-DETR模型对于小目标检出率达到92%,正检率达到95%,模型推理速度达16 f/s,满足高铁长大桥梁施工现场安全智能检测要求。
During the construction of long and large high speed railway bridges,the site scenes are complex and variable.To address the issue of poor accuracy in recognizing small targets such as construction personnel and safety helmets during high-altitude inspections with industry-grade UAV,this paper proposes an optimized intelligent detection method based on a real-time object detection model(RT-DETR).A lightweight Mobile-HGNet is designed to replace the backbone network of RT-DETR,effectively reducing the model’s computational load.Additionally,a cascaded grouped self-attention module is introduced,which not only ensures the model’s lightweight nature but also enhances both global and local region focus,thereby improving the detection performance of the model.Data augmentation is achieved using slicing aided hyper inference(SAHI)input,sacrificing a small amount of real-time performance in exchange for a higher detection rate.Comparing the improved RT-DETR model with the baseline model,the optimized RT-DETR model achieves a small target detection rate of 92%and a positive detection rate of 95%,with a model inference speed of 16 f/s.This meets the requirements for intelligent safety inspection at long and large high speed railway bridge construction sites.
作者
甘君
张崇斌
刘诗洋
周强
董晓兵
赵训刚
张宁军
GAN Jun;ZHANG Chongbin;LIU Shiyang;ZHOU Qiang;DONG Xiaobin;ZHAO Xungang;ZHANG Ning-jun(China Railway Major Bridge Engineering Group No.6 Engineering Co.,Ltd.,Wuhan Hubei 430050,China;China Railway Engineering Design and Consulting Group Co.,Ltd.,Beijing 100055,China;National Key Laboratory of Intelligent and Green Bridge Construction,Wuhan Hubei 430050,China;China Railway Major Bridge Engineering Group Co.,Ltd.,Wuhan Hubei 430050,China;China Railway Bridge Research Technology Co.,Ltd.,Wuhan Hubei 430050,China;2nd Engineering Co.,Ltd.of China Railway No.3 Engineering Group,Shijiazhuang Hebei 457000,China)
出处
《中国铁路》
北大核心
2025年第8期15-29,共15页
China Railway
基金
中国国家铁路集团有限公司科研开发计划项目(K2023G003)
中国中铁科研开发计划项目(2023-重大-02、2024-重点-07)
中铁工程设计咨询集团科研开发计划项目(研2023-5、KY2024A029、KY2024A004)。
关键词
无人机航拍
小目标智能检测
RT-DETR
模型改进
数据增强
UAV aerial photography
intelligent detection of small targets
RT-DETR
model improvement
data augmentation
作者简介
第一作者:甘君(1978-),男,高级工程师E-mail:798620583@qq.com。
