{
  "paper_id": "978afad79539ee604648e81cc41b4679",
  "paper_title": "EasyCalib: A Novel Target for High-Accuracy Fully Automatic Extrinsic Calibration of Camera and LiDAR",
  "paper_authors": [
    "Zhiyu Zhou",
    "Zhi Gao",
    "Min Cao",
    "Jingshi Wang",
    "Yong Li",
    "Zhipeng Lin",
    "Wenbin Yang",
    "Xiaonan Wang",
    "Qiyuan Zhang"
  ],
  "publication_date": "",
  "publication_venue": "IEEE Transactions on Instrumentation and Measurement",
  "abstract": "Precise extrinsic calibration is a fundamental requirement for enabling robust camera-light detection and ranging (LiDAR) fusion applications. Existing target-based methods suffer from practical limitations: they require manual annotation, precise target dimensions, and initial extrinsic estimates, demand complex target placement, and have limited applicability across different sensor types. To the best of our knowledge, we are not aware of a prior approach that simultaneously achieves full automation, avoids reliance on precise target dimensions and accurate initialization, and generalizes across both mechanical and nonrepetitive-scan solid-state LiDARs. In this article, we propose EasyCalib, a novel artificial target for high-accuracy, fully automatic extrinsic calibration for almost all types of camera-LiDAR systems. Specifically, we design dedicated point cloud segmenters for mechanical and solid-state LiDARs to automatically extract EasyCalib. By leveraging the co-visible corner points and central edge features of EasyCalib in both the image and point cloud domains, we achieve a coarse-to-fine, initialization-free, and highly accurate extrinsic calibration. Moreover, since the proposed calibration pipeline does not require the target to be strictly orthogonal and tolerates moderate assembly misalignment, and since the target dimensions are not involved in the calibration computation, EasyCalib requires neither complex manufacturing procedures nor precise prior knowledge of its size.",
  "keywords": [
    "Calibration target",
    "extrinsic calibration",
    "light detection and ranging (LiDAR) and camera"
  ],
  "paper_institutions": [
    "Wuhan University",
    "Hubei University of Technology",
    "Wuhan Guanggu Zoyon Science and Technology Company Ltd.",
    "Shanghai Jiao Tong University",
    "Shanghai Key Laboratory of Aerospace Intelligent Control Technology",
    "Shanghai Aerospace Control Technology Institute",
    "The Chinese University of Hong Kong",
    "ZG Technology Company Ltd."
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:42:37.817788Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n精确的相机与激光雷达（LiDAR）外参标定是鲁棒多传感器融合应用的基础需求，但现有基于标定物的方法存在需要人工标注、精确的标定物尺寸与初始外参估计、标定物摆放复杂以及跨传感器类型适用性有限等实际局限；无目标方法依赖深度不连续边缘、外观统计或独立运动等弱场景特征，对初始外参高度敏感，在光照差、纹理弱或里程计漂移时易失败，而平面/穿孔板需多姿态摆放与稠密点云，球体目标需高精度制造且尺寸误差会引入系统性偏差，多数现有目标也难以适配非均匀扫描的固态LiDAR。为此，本文提出EasyCalib——一种由红、绿、蓝三块扇形亚克力板构成的三面体标定物，并构建全自动由粗到细的标定流水线：针对机械式LiDAR设计自适应扫描线聚类（阈值 σ=2r·tan(θv/2)），针对非重复扫描固态LiDAR采用区域生长聚类，再经结构正则化（E_ang<τ_ang 且 E_dist<τ_dist）自动识别目标；图像端通过RGB三值化、Canny边缘检测与梯度方向直方图定位中心角点，点云端通过平面拟合与交线求解恢复三维角点，进而以RANSAC-PnP获得免初始化的粗外参，最后以点到线重投影残差 e_i = n_i^T(Π(^C T_L P_i) − q_i) 构建最小二乘并经高斯-牛顿优化实现亚像素级精化。实验在Gazebo仿真与真实场景中开展，涵盖Velodyne VLP-16、Livox Mid-360与Livox Avia三种LiDAR及三种针孔相机配置，以验证方法在机械式与非重复扫描两类扫描模式下的泛化能力；评价指标包括平移/旋转分量误差、真实场景重复性标准差、目标分割成功率与运行时间。结果显示，EasyCalib在多数自由度上取得最优精度（仿真中旋转误差低至0.01°量级），真实场景20次重复试验的旋转标准差低于0.1°、总运行时间小于1秒，且对三面体非正交偏差保持鲁棒，而对比的三面体方法显著退化，验证了免尺寸先验、免初始化与像素级精化设计的有效性。未来工作计划通过在面板集成加热元件将EasyCalib扩展至热红外相机标定，并利用中心线特征的3D线对应实现多LiDAR系统自动标定。实际部署中，该方法无需严格正交装配与复杂制造流程，仅需目标在LiDAR有效扫描范围内获得足够点云覆盖（实践采用38.8 cm扇形半径以支持15 m扫描范围），但严重遮挡或极稀疏采样（少于两条扫描线与目标平面相交）仍可能导致粗标定失败。\n\n### 关键要点\n\n- 目标设计：EasyCalib由三块红、绿、蓝扇形亚克力板构成三面体，其独特几何与RGB布局为LiDAR和相机数据提供可靠的检测与特征提取线索，且不要求严格正交装配。\n- 制造复杂度低：标定物尺寸不参与标定计算，无需精确几何或尺寸先验，避免制造误差引入系统性偏差；实践中扇形板半径设为38.8 cm，可在15 m的LiDAR扫描范围内正常工作。\n- 点云分割：针对机械式LiDAR提出自适应扫描线聚类（阈值 σ=2r·tan(θv/2) 随扫描距离与垂直角自适应调整），针对非重复扫描固态LiDAR采用区域生长空间聚类，实现目标自动定位。\n- 结构正则化：以平面法向量夹角一致性 E_ang 与平面中心距离一致性 E_dist 为判据（阈值 τ_ang=0.2、τ_dist=0.08 m）识别标定物簇；启用自适应σ时分割成功率在简单/困难场景分别达100%/92%，总体96%。\n- 特征提取：图像端采用RGB三值化、Canny边缘检测与梯度方向直方图（三主峰约120°均匀分布）检测中心角点；点云端通过平面拟合与两两平面交线求解三维中心角点（最小化到三条交线的平方欧氏距离之和）。\n- 粗-精标定机制：利用2D-3D角点对应通过RANSAC-PnP获得初始外参，再以点到线重投影残差 e_i = n_i^T(Π(^C T_L P_i) − q_i) 构建加权最小二乘，采用SE(3)李代数左扰动模型与解析雅可比，经高斯-牛顿迭代3–4次收敛至亚像素精度。\n- 实验设置：在Gazebo仿真与真实环境中评估三种传感器配置（Velodyne VLP-16+MER2630-60U3C相机、Livox Mid-360+MV-CU013相机、Livox Avia+D455相机），覆盖机械式与非重复扫描LiDAR以验证跨传感器泛化性。\n- 仿真精度：在三种传感器配置与三种目标摆放（单目标、同深度多目标、多深度多目标）下，EasyCalib在多数自由度上取得最佳精度（如Placement C下旋转误差0.01°–0.19°、平移误差1.17–3.21 cm），优于FastCalib、球体与三面体基线。\n- 真实场景重复性：20次重复试验中旋转标准差低于0.1°、平移标准差为亚厘米至数厘米；多深度目标摆放比单目标摆放更稳定，深度方向平移分量在约束不足时方差明显增大。\n- 鲁棒性验证：对三面体正交性偏差（90°±Δθ）不敏感，误差保持有界，而Choi等三面体方法随Δθ增大显著退化；点到线残差在3–4次高斯-牛顿迭代内快速收敛且无振荡。\n- 失败情形与约束：严重遮挡（中心角点三色分布不可见）或极稀疏LiDAR采样（少于两条扫描线与目标平面相交）会导致PnP粗标定失败；目标尺寸过小可能导致点云点数不足而无法提取平面特征。\n- 运行效率：在Intel i7-1360P CPU（16 GB RAM）上总运行时间为0.58–0.79秒，主要瓶颈为前端特征提取（点云聚类与图像处理），后端优化开销可忽略（粗标定0.1–0.2 ms，精标定亚毫秒级）。"
}