XiaoJake / PeLiCal

无目标的 RGBD 外参在线校准 [ICRA2024] Official Repository for PeLiCal: Targetless Extrinsic Calibration via Penetrating Lines for RGB-D Cameras with Limited Co-visibility

PeLiCal

Thanks to Dall-E for creating the mascot of PeLiCal!

This repository contains the code for PeLiCal, targetless extrinsic calibration for RGB-D cameras with limited co-visibility, which is accepted to ICRA 2024. PeLiCal leverages real-time line features from the environment for estimating the relative pose between cameras. To deal with the challenging configuration of cameras, we derive model-based approach for adding reliable feature and rejecting outliers. Also, we robustly estimate pose by merging two types of constraints into quadratic system.

[arXiv] [BibTex]

News

• 24.04.22 We released an inital version of PeLiCal!

1. Dependencies

The code requires following dependencies:

• Ubuntu 20.04 and ROS Noetic
• Ceres Solver (2.1.0)
• Eigen3 and Sophus
• GlueStick
• OpenCV4 (4.5.2)
• PoseLib (2.0.0)
• RealSense-ROS1

2. Install

cd ~/[your ROS Workspace]/src
git clone https://github.com/joomeok/PeLiCal.git
cd ..
catkin build
source devel/setup.bash

3. Launch PeLiCal in ROS

3.1 Create .yaml file

Create your own Calibrator.yaml file containing principal information for the cameras under config directory. For example:

image_width: 640
image_height: 480
source_rgb: "/cam_1/color/image_raw"
source_depth: "/cam_1/aligned_depth_to_color/image_raw"
target_rgb: "/cam_2/color/image_raw"
target_depth: "/cam_2/aligned_depth_to_color/image_raw"

voting_eps: 7.0
voting_thresh: 6
minimum_line_num: 15
convergency_cost: 10.0

target_intrinsic: 
  data: [599.679256, 0.000000, 320.187940, 0.000000, 600.098148, 244.389703,0,0,1]

source_intrinsic:
  data: [598.330788, 0.000000, 316.565022,0.000000, 598.405846, 232.152461,0,0,1]

gluestick_root: "/home/jaeho/GlueStick/"

Source camera refers to the one positioned at the right side based on the viewing direction, and target camera is positioned at the left side.

3.2 Run PeLiCal

First, launch your own data (aligned depth image from RealSense-ROS), for example:

roslaunch realsense2_camera rs_camera.launch align_depth:=true camera:=cam_1 serial_no:=<serial number of the first camera>
roslaunch realsense2_camera rs_camera.launch align_depth:=true camera:=cam_2 serial_no:=<serial number of the second camera>

Then, open another terminal and execute following command:

roslaunch pelical calibration.launch

3.3 Additional Parameters

We offer 4 parameters for tuning the performance of the algorithm in accordance with the data.

• voting_eps: Counting threshold for convergence
• voting_thresh: Convergency threshold for translation candidates
• minimum_line_num: Minimum number of lines for terminating the algorithm
• convergency_cost: Convergency thershold for terminating the algorithm

Demo Video

Here are simple demo video for calibrating two RealSense D435i cameras with 60 degree apart.

BibTex

@INPROCEEDINGS{jhshin-2024-icra,  
    AUTHOR = { Jaeho Shin and Seungsang Yun and Ayoung Kim },  
    TITLE = { PeLiCal: Targetless Extrinsic Calibration via Penetrating
Lines for RGB-D Cameras with Limited Co-visibility },  
    BOOKTITLE = { IEEE International Conference on Robotics and Automation (ICRA) },  
    YEAR = { 2024 },  
    MONTH = { May },  
    ADDRESS = { Yokohama },  
}