Official Implementation of the paper Targetless Lidar-camera Calibration via Cross-modality Structure Consistency

Paper

• Techrxiv
• IEEE

KITTI Trajectory

00	02	03

04	05	07

These trajectories are not suitable for hand-eye calibration because of lack of adequent rotations around at least two directions (only in yaw). Other hybrid methods may also suffer from the bad initial extrinsic estimation from hand-eye calibration. We believe this proposed method offers a valuable tool to resolve this problem.

Performance on KITTI Odometry

• rotation units: degree
• translation units: cm
• Accuracy reserved to 0.01

• Examples of Quanlitative Results

Seq	Predicted	Ground-truth
00
02
03
04
05
07

Tested Environment

More recent version would be OK.

Dependencies

• Eigen 3 (libeigen3-dev)
• OpenMP (libomp-dev)
• OpenCV 4.x
• Pangolin (visual tools for ORB_SLAM2)
• g2o 2023 (-DG2O_USE_OPENMP=ON)
• Open3D (Point Cloud IO, ICP)
• Nomad (Global Optimization Algorithm)

Install

• Copy Thirdparty and Vocabulary directories to src/orb_slam/
  Troubleshooting

  Note that the implementation of ORB_SLAM2 in our repo is different from the original one, so DO NOT copy the whole ORB_SLAM2 repo to replace our directory
• Build and Compile: cd build && cmake .. -Wno-dev && make -j
  TroubleShooting

  If you have installed g2o through ROS (if you have ROS packages like `base_local_planner`/`teb_local_planner`/`mpc_local_planner`), please exclude it from LD_LIBRARY_PATH environment variable, or `source config/settings.sh`.
• Argparse Support (Acknowledge argparse ) : use -h or --help for each executable cpp file to view help.

Step 1: Estimate Camera and Lidar Poses

Why SLAM?

Our algorithm requires 3D visual map (scaleless), 3D pointcloud map, 2D feature points for each frame, relative poses of cameras and LiDAR.

orb_slam

An example command for KITTI 00 Sequence:

./orb_store ../data/00/ ../config/orb_ori/KITTI00-02.yaml ../data/Vocabulary/ORBvoc.txt ../KITTI-00/slam_res/Twc.txt ../KITTI-00/KeyFrames/ ../KITTI-00/Map.yml --slow_rate 1.5

• ../data/00/: sequence directory of KITTI Sequence 00
• ../config/orb_ori/KITTI00-02.yaml: yaml file for ORB_SLAM
• ../data/Vocabulary/ORBvoc.txt: DBoW2 Vocabulary txt file for ORB_SLAM
• ../KITTI-00/slam_res/Twc.txt: KeyFrame poses saved by ORB_SLAM, in KITTI format
• ../KITTI-00/KeyFrames/: KeyFrame information saved by ORB_SLAM, can be restored during runtime
• ../KITTI-00/Map.yml: Map saved by ORB_SLAM, can be restored during runtime
• --slow_rate 1.5 [Optional]: Slow Rate of ORB_SLAM2. Duration between Frames: computation + wait_time >= 1.5 * real timestamp. Larger slow_rate is needed for a less powerful CPU. 1.5 is recommended for a I9-12900KF CPU.

F-LOAM

an example command for KITTI 00 Sequence:

./floam_run ../data/00/velodyne/ ../KITTI-00/slam_res/floam_raw_00.txt

• ../data/00/velodyne/: directory containing the Lidar PointClouds, whose filenames must be sorted by timestamp.
• ../KITTI-00/slam_res/floam_raw_00.txt: Lidar Poses Estimated by F-LOAM.

F-LOAM Backend Optimization

An example command for KITTI 00 Sequence:

 ./floam_backend ../config/loam/backend.yml ../KITTI-00/slam_res/floam_raw_00.txt ../data/00/velodyne/

• ../config/loam/backend.yml: config file of backend optimzation
• ../KITTI-00/slam_res/floam_raw_00.txt: Lidar poses estimated by F-LOAM
• ../data/00/velodyne/: directory containing the Lidar PointClouds, whose filenames must be sorted by timestamp.

Expected Results

• ../KITTI-00/KeyFrames: KeyFrame directory includes information of KeyFrames
• ../KITTI-00/Map.yml: Visual Map built by ORB_SLAM
• ../KITTI-00/FrameId.yml: Yaml file that contains Frame indicces of KeyFrames (will saved in the same directory of "Map.yml"
• ../KITTI-00/slam_res/Twc.txt: KeyFrame Poses estimated by ORB_SLAM
• ../KITTI-00/slam_res/floam_raw_00.txt: Lidar Poses estimated by F-LOAM (the number of itmes in this file is not equal to that in Twc.txt)
• ../KITTI-00/slam_res/floam_isam_00.txt: Lidar Poses Optimized by iSAM ((the number of itmes in this file is not equal to that in Twc.txt but equal to floam_raw_00.txt

Step 2: Hand-eye Calibration with Regularization

An example command for KITTI 00 Sequence:

 ./he_calib ../config/calib/00/he_calib.yml

• A file with 13 entries will be saved to ../KITTI/calib_res/he_rb_calib_00.txt.
• The calibration result of the orindary hand-eye calibration is also saved for checking: ../KITTI/calib_res/he_calib_00.txt.

Step 3: Global optimization

Nomad library must be correctly installed before this step.

An example command for KITTI 00 Sequence:

 ./iba_global ../config/calib/00/iba_calib_global.yml

The final calibration result will be saved to ../KITTI-00/calib_res/iba_global_pl_00.txt Two parameters in ../config/calib/00/iba_calib_global.yml are used to create variants for ablation experiments:

runtime:
  err_weight: [1.0, 1.0]
  use_plane: true

• set use_plane to true and err_weight to [1.0,1.0] to apply CBA+CA (PT+PL) method in Table 1 of our paper (proposed).
• set use_plane to false and err_weight to [1.0,1.0] to apply CBA+CA (PT) method in Table 1 of our paper.
• set use_plane to false and err_weight to [1.0,0.0] to apply CBA method in Table 1 of our paper.

Remember to set different

io:
  ResFile: calib_res/iba_global_baonly_00.txt

for different methods, or the prior files will be overwritten.

Parameters that need to be changed in other sequences

• ../data/00/ -> ../data/xx/ for the orb_store program;
• ../config/orb_ori/KITTI00-02.yaml -> ../config/orb_ori/KITTIxx-xx.yaml;
• KITTI-00 -> KITTI-xx for all aurgments;
• ../data/00/velodyne/ -> ../data/xx/velodyne/;
• floam_raw_00.txt -> floam_raw_xx.txt;
• floam_isam_00.txt -> floam_isam_xx.txt;