浅埋连拱隧道施工场景风险行为视觉关系检测仿真  

Visual Relationship Detection Simulation of Risk Behavior in Shallow Buried Multi-arch Tunnel Construction Scene

作　　者：马培广 邓稀肥 MA Peiguang;DENG Xifei(China Railway(Guangzhou)Investment&Development Co.Ltd.,Guangzhou 510335,China;China the Si Ju Civil Engineering Group,Hefei 230023,China)

机构地区：[1]中铁广州投资发展有限公司,广州510335 [2]中铁四局集团有限公司,合肥230023

出　　处：《武汉理工大学学报（交通科学与工程版）》2023年第5期928-932,共5页Journal of Wuhan University of Technology（Transportation Science & Engineering）

摘　　要：文中提出浅埋连拱隧道施工场景风险行为视觉关系检测仿真方法.基于对抗网络修复监控视频图像,避免发生伪影现象,提升图像的清晰度和质量.将修复后的图像输入YOLOv3检测模型中,通过模型中的残差模块和上采样提取并融合图像特征,生成特征图.在模型的残差层和特征层之间,引入通道域注意力机制,优化特征提取结果,强化特征图,显著降低背景信息的干扰,获取更丰富的特征信息,输出风险行为检测结果.结果表明:该方法的图像修复效果良好,精准判断施工场景内是否存在检测目标,前景和背景误检率高值分别为4.5%和5.5%,可实时、可靠地完成施工场景内各种风险行为的检测.A simulation method of visual relationship detection of risk behavior in shallow-buried multi-arch tunnel construction scene was proposed.Based on the countermeasure network,the surveillance video image was repaired to avoid artifacts and improve the clarity and quality of the image.The restored image was input into YOLOv3 detection model,and image features were extracted and fused by residual module and up-sampling in the model to generate feature map.Between the residual layer and the feature layer of the model,the channel domain attention mechanism was introduced to optimize the feature extraction results and strengthen the feature map,which significantly reduced the interference of background information.More abundant characteristic information was obtained,and the risk behavior detection result was output.The results show that the image inpainting effect of this method is good,and it can accurately judge whether there is a detection target in the construction scene.The high false detection rates of foreground and background are 4.5%and 5.5%respectively,and it can reliably detect various risk behaviors in the construction scene in real time.

关 键 词：浅埋连拱隧道 施工场景 风险行为 视觉关系检测 图像修复 特征图强化 

分 类 号：U455[建筑科学—桥梁与隧道工程]