基于非接触式机器视觉监测技术与区间仿射算法的桥梁结构影响面反演

As a novel technology in the structural health monitoring of bridge, non-contact machine vision has obvious advantages including rapid implementation, easy operation, and global measurement, and it provides important significance in developing machine vision identification technology for the rapid bridge condition assessment. However, this technology is significantly affected by the vehicle weight extraction relying on the support of bridge weight-in-motion (BWIM) system, the inaccurate spatial-temporal location of vehicles, and the physical target attachment for deflection measurement. Therefore, this research project aims to investigate the inversion of bridge influence surface based on non-contact machine vision monitoring technology and interval affine arithmetic, combining the advanced methodologies of deep-learning theory, computer graphics, and interval algorithm. The main study focuses on vehicle feature extraction and wheel load estimation based on the deep learning theory, estimation of vehicle spatial-temporal location based on three-dimensional coordinate transformation in the parallel coordinate system, identification of bridge dynamic deflection based on the fast retina key-point (FREAK) technique, and inversion of bridge deflection surface based on the interval affine arithmetic. The research results of this project are expected to broaden the area of non-sensor installation on the small and middle span bridges for the in-service condition assessment, provide a new ideology in the rapid bridge condition assessment without traffic closure, sensor installation and BWIM system support, and offer a new method to realize the intelligent bridge and intelligent transportation.

非接触式机器视觉识别技术在桥梁健康监测中具有“快速实施”、“操作简便”及“全局测量”等明显优势，开发能进行桥梁状态快速评估的机器视觉识别系统具有重要的意义。然而，由于存在着车辆载荷提取依赖桥梁动态称重(BWIM)系统、车辆时空定位精度不高和变形测试依赖物理标靶等方面的不足，本项目将综合深度学习理论、计算机图形学和区间算法等领域的先进方法，开展基于非接触式机器视觉监测技术与区间仿射算法的桥梁结构影响面反演研究。主要研究基于深度学习理论的车辆类型识别及轮载区间估计，基于平行坐标系三维空间坐标转换的车辆时空定位，基于快速视网膜关键点追踪的桥梁结构动位移识别和基于区间矩阵仿射算法求逆的桥梁影响面反演。本项目的开展将拓宽无传感器安装的中小桥梁长期运营监测的范围，提供了一种“无需交通封闭”、“无需接触式传感”和“无需BWIM系统支持”的桥梁快速评估的新思路，是实现“智能桥梁”和“智慧交通”的重要手段。

桥梁健康监测中机器视觉识别方法存在车辆载荷提取依赖桥梁动态称重系统、车辆时空定位精度不高和变形测试依赖物理标靶等不足，本项目开发出了基于机器视觉技术识别车辆荷载和桥梁变形的核心算法，实现了非接触式桥梁快速检测和桥梁结构力学性能评估。.(1)提出了一种基于桥面视频的车辆自动检测跟踪分类系统。利用深度学习算法Faster R-CNN对车辆位置进行检测，通过卡尔曼滤波跟踪算法对多目标进行跟踪，结合深度卷积神经网络VGG-16和车辆数据信息库获得了桥上车辆的基本类型、轴重区间和实时轨迹。.(2)提出了一种基于深度学习理论的车辆荷载辨识方法。利用卷积神经网络代替结构系统建模，实现了仅利用结构特征响应输出获取输入信息。针对某野外桥梁试验获得了97.21%的测试精度，验证了该技术识别车辆荷载的可靠性。.(3)提出了一种基于摄影测量的车辆时空定位技术。以运营车辆作为参照物，利用级联霍夫变换方法对正交消失点进行检测，完成了无需标靶的相机自标定。利用相机模型对车辆荷载在桥面上的真实位置进行了识别，平均误差约为1.75%。.(4)提出了一种基于数字图像相关理论的非接触式结构位移测量方法。采用改进初值的亚像素算法确定待测结构目标位置的基于像素的位移，并利用相机标定方法转换为工程单位，钢框架结构模型的试验表明该方法的识别精度达毫米级。.(5)提出了一种基于区间仿射算法的非接触桥梁结构影响线识别方法。基于含有不确定输入参数的区间分析理论，结合仿射矩阵求逆算法从实测桥梁响应中反演出影响线区间，利用SVM算法提取真实影响线，识别精度达92.57%。

内容获取失败，请点击重试