大跨桥梁车辆追踪与荷载时空分布智能识别

Intelligent Identification of Vehicle Tracking and Load Spatio-temporal Distribution in Long-span Bridge

车辆荷载是大跨桥梁最重要的作用荷载之一,也是大部分桥梁疲劳劣化的最主要原因。但桥梁动态称重系统造价昂贵,无法在桥上分布式布置,桥梁车辆荷载分布信息的动态识别仍是挑战性难题。面向大跨桥梁结构健康监测需求,引入计算机视觉与深度学习技术,建立了一套集成化的桥梁车辆荷载时空分布智能识别系统。首先,研究基于交通监控数据和深度目标检测网络的车辆识别方法,开展车辆目标检测任务的YOLOv7深度网络训练,并通过训练后模型获取单摄像头数据中包含车辆类型与时间等信息的车辆图像;其次,引入HardNet深度特征描述符,建立图像点特征匹配方法,通过分布布置的监控视频数据设计搜索匹配策略,实现车流方向多个监控对应车辆图像数据的匹配,并对监控盲区采用线性插值估计车辆位置,得到车辆在桥梁上的时空分布信息;然后,将各方法集成,建立车辆荷载时空分布识别系统,该系统可结合动态称重数据自动输出车辆荷载时空分布信息与可视化结果,实现从监控数据到车辆荷载时空分布的一体化流程;最后,采用九江长江大桥监控数据进行应用验证。研究结果表明:该系统基于视频数据实现车辆目标识别与匹配追踪,运算耗时小于输入视频时长,对大型车辆匹配准确率达97.62%,可以快速、准确地识别车辆荷载分布信息,该系统对保障桥梁服役安全具有重要的意义,应用前景广阔。

Vehicle load is one of the most important loads of long-span bridges,and it is also the main cause of fatigue deterioration of most bridges.However,the bridge weigh-in-motion system is expensive and cannot be distributed across the bridge,which means the dynamic identification of bridge vehicle load distribution information is still a challenging problem.This paper introduced computer vision and deep learning technologies to meet the needs of long-span bridge structural health monitoring,and established an integrated intelligent identification system for bridge vehicle load spatio-temporal distribution.Firstly,we studied the vehicle identification method based on traffic monitoring data and deep target detection network,trained the YOLOv7 deep network for vehicle target detection tasks,and obtained vehicle images containing information such as vehicle type and time in single camera through the trained model.Then,we introduced the HardNet depth feature descriptor to establish an image point feature matching method,designed a searching and matching strategy through distributed surveillance video data to achieve the matches of vehicle image data corresponding to multiple monitors in the traffic flow direction,and the vehicle position was estimated by linear interpolation of the monitoring blind area to obtain the spatio-temporal distribution of vehicles on the bridge.Finally,the methods were integrated to establish the vehicle load spatio-temporal distribution identification system.This system can automatically output the spatio-temporal distribution of vehicle load and visualization results combining with dynamic weighing data,realizing an integrated process from monitoring data to vehicle load spatio-temporal distribution.In this paper,the monitoring data of Jiujiang Yangtze River Bridge was used for verification.The results show that the system can achieve vehicle identification and tracking based on video data,with computational time less than the duration of the input video and an accuracy rate of 97.62%for large vehicle matching,allowing for rapid and accurate identification of vehicle load distribution.The system is of great significance to ensure the safety of bridge service and has broad application prospect.