{
  "paper_id": "841dc43eccd867de58f0034a4510a37c",
  "paper_title": "Bridge Geometric Shape Measurement Using LiDAR-Camera Fusion Mapping and Learning-Based Segmentation Method",
  "paper_authors": [
    "Shang Jiang",
    "Yifan Yang",
    "Siyang Gu",
    "Jiahui Li",
    "Yingyan Hou"
  ],
  "publication_date": "2025-04-25",
  "publication_venue": "Buildings (MDPI), 2025, Vol. 15, Article 1458",
  "abstract": "The rapid measurement of three-dimensional bridge geometric shapes is crucial for assessing construction quality and in-service structural conditions. Existing geometric shape measurement methods predominantly rely on traditional surveying instruments, which suffer from low efficiency and are limited to sparse point sampling. This study proposes a novel framework that utilizes an airborne LiDAR–camera fusion system for data acquisition, reconstructs high-precision 3D bridge models through real-time mapping, and automatically extracts structural geometric shapes using deep learning. The main contributions include: (1) A synchronized LiDAR–camera fusion system integrated with a UAV and a microprocessor was developed, enabling flexible and large-scale acquisition of bridge images and point clouds; (2) A multi-sensor fusion mapping method coupling visual-inertial odometry (VIO) and LiDAR-inertial odometry (LIO) was implemented to construct 3D bridge point clouds in real time robustly; and (3) An instance segmentation network-based approach was proposed to detect key structural components in images, with detected geometric shapes projected from image coordinates to 3D space using LiDAR–camera calibration parameters. The proposed method was validated through geometric shape measurements on a concrete arch bridge. The results demonstrate that compared to the oblique photogrammetry method, the proposed approach reduces errors by 77.13%, while its detection time accounts for 4.18% of that required by a stationary laser scanner and 0.29% of that needed for oblique photogrammetry.",
  "keywords": [
    "bridge geometric shape",
    "LiDAR",
    "point cloud",
    "3D reconstruction",
    "deep learning"
  ],
  "paper_institutions": [
    "School of Transportation and Civil Engineering, Nantong University, Nantong 226019, China",
    "School of Civil Engineering, Southeast University, Nanjing 211189, China",
    "Target Key Laboratory of Cognition and Application Technology, Aerospace Information Research Institute, Beijing 100190, China"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:31:00.452230Z",
  "template_tag": "simple",
  "summary": "### 摘要\n桥梁三维几何形状的快速测量对评估施工质量与在役结构状态至关重要，但现有手段存在明显不足：全站仪逐点测量需人工预铺反射棱镜、危险且耗时、采样稀疏；静态地面激光扫描（TLS）对长桥需布设30个以上测站，外业与配准动辄数小时；无人机倾斜摄影精度受相机质量、光照、阴影遮挡与影像重叠率影响，大规模重建难以达到厘米级且算法极为耗时；而基于深度学习的大规模点云分割又受制于真实数据稀缺（多依赖合成数据集）与GPU算力需求（处理不足140万点的桥梁点云也需30分钟以上）。针对上述问题，本研究提出一套机载LiDAR-相机融合测量框架：将Livox Avia非重复扫描LiDAR（扫描范围超190m、24万点/秒、内置BMI088 IMU）、HIKROBOT MV-CE120-10UC工业相机（4000×3036像素）与Intel NUC12处理器集成于DJI M300 RTK无人机，通过STM32硬件触发（LiDAR 1Hz、相机10Hz PWM信号）将传感器同步误差控制在约10ms；采用耦合视觉惯性里程计（VIO）与激光惯性里程计（LIO）的FAST-LIVO紧耦合误差状态迭代卡尔曼滤波（ESIKF）框架实时重建三维点云，其中利用IMU预积分运动方程（如 P_{t+Δt}=P_t+V_tΔt+½gΔt²+½R_t(â_t−b_t^a−n_t^a)Δt²）消除动态扫描引起的点云运动畸变；进而提出基于Transformer与四叉树层级结构的实例分割网络，先在二维图像中精细分割桥梁构件边缘，再依据LiDAR-相机标定参数（ACSC方法，平移误差2.61mm、重投影误差2.48像素）将分割掩膜投影回三维点云提取构件坐标，从而规避大规模点云的直接处理。方法在一座建于1970年代、主跨140m、桥宽8m的混凝土拱桥上验证：分割网络基于300张不同光照条件下人工标注的无人机拱圈图像训练（RTX 3090 GPU、PyTorch 1.13.0、1000次迭代、学习率0.001），测试精度达94.23%，单幅4000×3000像素图像分割约0.5秒，图像分割到点云分割全流程平均2秒；以高精度TLS（RIEGL VZ-400i）为真值、80个反射靶标评估，本方法X/Y/Z轴平均误差分别为0.083m、0.103m、0.061m，较倾斜摄影（0.452m、0.602m、0.184m）平均误差降低77.13%，总耗时（0.167h）仅为TLS工作流（4h）的4.18%、倾斜摄影（57.67h）的0.29%，且因扫描距离一致（约15m）保持全桥点云密度均匀。未来可通过选用更高精度LiDAR进一步提升精度，并扩充拉索、箱梁、桥塔等构件数据集以推广至更多桥梁类型；当前约束包括低成本LiDAR硬件精度限制（相对静态TLS存在8.3cm误差）、分割网络仅针对混凝土拱桥拱圈、以及方法仅适用于本融合系统采集的数据。\n\n### 关键要点\n- 系统构成：Livox Avia LiDAR（190m以上扫描范围、24万点/秒、集成0.09mg分辨率BMI088 IMU）与HIKROBOT MV-CE120-10UC工业相机（5mm焦距镜头、4000×3036像素）经3D打印框架连接，配合Intel NUC12（i5-12500H、16GB RAM）处理器，搭载续航约35分钟的DJI M300 RTK无人机，实现灵活的大规模桥梁图像与点云采集。\n- 时间同步：STM32微控制器生成1Hz PWM方波触发LiDAR、10Hz PWM信号触发相机，将传感器间同步误差降至约10ms，克服软件触发方式因时间抖动导致的累积失配问题。\n- 标定流程：Zhang法标定相机内参，基于IMU预积分的LiDAR-IMU标定消除动态扫描运动畸变，ACSC方法（时空几何特征提取结合强度分布分析）估计LiDAR-相机外参，标定验证平移误差2.61mm、重投影误差2.48像素。\n- 3D重建框架：选用FAST-LIVO紧耦合ESIKF融合LiDAR、相机与IMU数据，其VIO子系统复用LiDAR点云并绕过视觉特征提取、三角化与优化，在算力受限和传感器退化环境下保持鲁棒，运行效率与建图稳定性优于R3LIVE++和LVI-SAM。\n- 分割网络：基于Transformer的实例分割网络聚焦构件边缘精细识别，以四叉树结构（最大深度3）表示不同RoI层级（28×28至56×56）的信息损失节点，包含节点编码器、序列编码器与像素解码器，通过多头注意力实现多尺度稀疏特征的长程比较与跨层级关联。\n- 数据与训练：300张不同光照条件下无人机拍摄的拱圈图像经人工标注构成训练评估集；训练硬件为i7-11700K CPU、32GB RAM与RTX 3090 GPU，采用PyTorch 1.13.0框架，1000次迭代、学习率0.001，最终测试精度94.23%。\n- 处理效率：单幅4000×3000像素图像的自动分割约需0.5秒；从图像拱圈分割到对应点云分割（点云帧积分时间100ms）的总处理时间平均为2秒。\n- 现场测试对象：1970年代建造的混凝土拱桥，主跨140m、桥宽8m，服役数十年后出现拱圈变形与侧倾问题，需精确三维几何检测以评估结构状态。\n- 三方法对比：TLS（RIEGL VZ-400i，14个测站、40个反射靶标、总耗时4h、8.82亿原始点）；倾斜摄影（DJI P1相机2068张图像、COLMAP处理共57.67h、2.55亿点）；本文方法（两次飞行共10分钟、1874万原始点、去除背景后1207万点）。\n- 精度结果：以TLS提取的80个反射靶标坐标为真值，本方法X轴平均误差0.083m（为倾斜摄影0.452m的18.36%）、Y轴0.103m（0.602m的17.11%）、Z轴0.061m（0.184m的33.15%），平均误差较倾斜摄影降低77.13%。\n- 效率与密度优势：总耗时仅为TLS工作流的4.18%、倾斜摄影的0.29%；因无人机以约15m一致距离绕桥扫描，全桥点云密度均匀，而TLS在跨中区域因扫描距离过远导致点云稀疏。\n- 局限与展望：低成本非重复扫描LiDAR硬件精度低于高精度LiDAR，导致相对静态TLS存在8.3cm重建误差；分割网络目前仅针对混凝土拱桥拱圈；方法依赖图像-点云标定投影，仅适用于本融合系统采集的数据，后续将扩充拉索、箱梁、桥塔数据集以推广至更多桥梁类型与构件。"
}