rayleizhu / plcc

Plane constraint based Lidar-Camera Calibration

PLCC

Plane constraint based Lidar-Camera Calibration.

1. Quick Start

We assume you have setup ROS enviroment.

1.1 Installation

WARNING: the master branch is in development, please checkout v1.0 branch to get stable and usable code.

git clone https://github.com/rayleizhu/plcc.git
cd plcc
git checkout v1.0
bash install.bash
source devel/setup.bash

The installation script is tested on Ubuntu 18.04. If you fail to install some python library, please install it mannually. See requirements.txt.

1.2 Test with sample data

We provided sample datadata/input/data.bag, you can run a test by executing the following command

cd scripts
bash run_all.bash

When the script terminated, you can get results in the directory data/output/. Two import files are:

• calib_result.csv: containing calibration result (R, t) which is transformation from lidar to camera
• reprojection.png: the reprojection result using calibrated realative transformation.

Data directory structure after calibration:

Example of reprojected result:

2. Run calibration with your own data

2.1 Procedures

1. Replace the sample data data/input/data.bag with your own recorded bag file. Refer to part 2.2 for requirement of the ROS bag file.
2. Modify configurations in scripts/run_all.sh
3. Run the following commands

cd scripts
bash run_all.bash

2.2 Requirement

You can use rosbag info data/input/data.bag to figure out how the input bag should be like:

Topics:

• /tf: should contain transformations from each tag to camera. You can use apriltag_ros to achieve this goal.
• /velodyne_points: the topic containing point cloud. The name of this topic is depend on the sensor you are using. Importantly, you need to change PCD_FRAME in run_all.bash corrspondingly as the frame point cloud resides in.
• /zed/zed_node/left/image_rect_color: the topic containing images camera shot. Corresponding variable in run_all.bash is IMG_TOPIC.
• /zed/zed_node/left/camera_info: the topic containing images camera information (intrinsics, etc.). Corresponding variable in run_all.bash is CAM_TOPIC.

Note that, When you record data, we assume the x axis of LiDAR points to calibration target planes, we use this assumption to figure out correct direction of plane normal (each plane has two-side normals), see line 197 of fit_plane_in_pcd.py.

2.3 Tips

1. you'd better

3. How to use our rviz plugin for point cloud patch selection

hold ALT: moving mode
hold SHIFT: incremental selection
hold CTRL: inverse selection
press F: focus on selected points
press P: pushlish selected patch

4. How it works

The main idea is to use planes to construct correspondence constraints: for each plane we can derive its equation in both camera frame (e.g. using Apriltag detection or Chessboard) and LiDAR frame (fitting point cloud patch), then the normal can be used to constraint roation, and intersection of multiple planes can be used to constraint translation.

5. Explanation of files

During initial development, I deliberately partition the tool into several modules, many of them can be run seperately. This is for convenience of debugging and readability. It's easy to understand each module by the file name, here I briefly introduce what each file does:

5.1 Core scripts

• preprocess.py: it helps you extract image captured by camera, camera-tag transformation, camera info, etc. from rosbag file.
• fit_plane_in_pcd.py: fits plane equation from point cloud patches.
• fina_plane_in_cam.py: get plane equation from cam-tag transformation.
• solve_tf.py: solve Lidar-Camera extrinsic based on plane constraints.
• project_pcd_to_img.py: for reprojection.
  The above scripts can be run seperately (you can run any of them with -h option to see its arguments) for calibration, or be imported for further development.

5.2 Other scripts

• run_all.sh: it connect the whole pipeline with bash script, so that you can run calibration with one command.
• camera_model.py: as its name implies, it includes camera model implementation. Currently, pinhole camera and MEI camera are provided.
• utility.py: some commonly used functions are put into this file.
• generate_target.py: you can generate calibration target (April tag or ChArucoBoard) with it.
• diff_extr.py: solve relative transformation. It's useful to validate calibration precision in a 'differential' manner, e.g., let's say you know the extrinsics between left eye and right eye of ZED stereo camera, and you derive both LiDAR-left extrinsic and LiDAR-right extrinsic with our tool, then you can get 'measured' left-right extrinsic with diff_extr.py, finally, you may validate the predcision by comparing it with groud truth extrinsic.

6. Related projects

selected_points_publisher: We refer to this repo to write our rviz plugin for point cloud patch selection.
camodocal: We use this repo to calibrate camera intrinsics.
apriltag_ros: We use it to get camera-tag transformation.

7. TODO

• Refactor ChArucoBoard detection part
• Refactor the code so that we can run the calibration with one python script and one config file
• Part 2.3 tips
• Finish the transformation calculator GUI tool