How hku-mars/livox_camera_calib works
Project Type and Purpose
• This is a ROS-based C++ toolkit (“lidar_camera_calib”) for extrinsic calibration between a high-resolution LiDAR sensor and a camera, in both indoor and outdoor, targetless environments. It provides single-scene and multi-scene calibration modes, producing pixel-level alignment accuracy by leveraging PCL, Eigen, and Ceres Solver.
Key Architectural Components
a. ROS Infrastructure
• ROS Master & Parameter Server
• Topics:
– /camera/image_raw (sensor_msgs/Image via cv_bridge)
– /velodyne_points or /pcd_cloud (sensor_msgs/PointCloud2)
b. Nodes
• bag_to_pcd node (src/bag_to_pcd.cpp)
– Reads ROS bags or live LiDAR streams
– Converts to PCL PointCloud2 and writes .pcd files
• Single-scene calibration node (src/lidar_camera_calib.cpp)
– Loads YAML config (config/calib.yaml) via rosparam
– Subscribes to image + point cloud topics
– Extracts edges, formulates optimization problem
– Uses Eigen for transforms + Ceres for non-linear optimization
– Publishes calibrated extrinsic to text (result/extrinsic.txt)
– Publishes debug/visualization topics for RViz
• Multi-scene calibration node (src/lidar_camera_multi_calib.cpp)
– Similar pipeline, loops over N scene pairs from config/multi_calib.yaml
– Rough calibration followed by fine optimization
c. Configuration & Assets
• YAML files in config/ (single and multi-scene parameters, file paths, camera intrinsics)
• RViz configurations in rviz_cfg/ for visualizing point cloud overlaid on image
• Example data & results in pics/ and result/
d. External Libraries & Build
• PCL (Point Cloud Library)
• Eigen (linear algebra)
• Ceres Solver (optimization)
• ROS packages: cv_bridge, pcl_conversions
• Built with catkin/CMake (CMakeLists.txt, package.xml)
Relationships and Data Flow
1. Launch File (launch/calib.launch or multi_calib.launch)
→ Loads launch arguments & rosparams from config/*.yaml
→ Starts bag_to_pcd (if using recorded data)
→ Starts calibration node(s)
2. bag_to_pcd node reads LiDAR stream → writes .pcd & publishes PointCloud2
3. Calibration node:
– Subscribes PointCloud2 + Image
– Edge extraction → Match 2D image edges with 3D cloud edges
– Ceres optimization → Compute extrinsic transform T_lidar_to_cam
– Publishes transform and viz topics → rviz for overlay
– Writes final extrinsic.txt
4. Multi-scene node loops above steps over multiple input pairs, aggregates results.
Architectural Patterns & Principles
• Modularity via ROS nodes
• Separation of concerns: data acquisition (bag_to_pcd), processing/optimization, visualization
• Configuration-driven via YAML + ROS parameter server
• Pipeline pattern: sensor data → feature extraction → optimization → result publication
• Use of standard robotics middleware (ROS topics/services)