Home

Awesome

LiDARTag

Overview

This is a package for LiDARTag, described in paper: LiDARTag: A Real-Time Fiducial Tag System for Point Clouds (PDF)(arXiv). This work is accepted by IEEE Robotics and Automation Letters and published at (here).

Image-based fiducial markers are useful in problems such as object tracking in cluttered or textureless environments, camera (and multi-sensor) calibration tasks, and vision-based simultaneous localization and mapping (SLAM). However, the state-of-the-art fiducial marker detection algorithms rely on the consistency of the ambient lighting. To the best of our knowledge, there are no existing fiducial markers for point clouds.

This paper introduces LiDARTag, a novel fiducial tag design and detection algorithm suitable for LiDAR point clouds. The proposed method runs in real-time and can process data at 100 Hz, which is faster than the currently available LiDAR sensor frequencies. Additionally, the software works with different marker sizes in cluttered indoors and spacious outdoors, even when it is entirely dark. Because of the LiDAR sensors' nature, rapidly changing ambient lighting will not affect the detection of a LiDARTag. Therefore, LiDARTag can be used in tandem with camera-based markers to address the issue of images being sensitive to ambient lighting.

This package has been tested under [ROS] Melodic and Ubuntu 18.04.
[Note] More detailed introduction will be updated shortly. Sorry for the inconvenient!
[Issues] If you encounter any issues, I would be happy to help. If you cannot find a related one in the existing issues, please open a new one. I will try my best to help!

Abstract

Image-based fiducial markers are useful in problems such as object tracking in cluttered or textureless environments, camera (and multi-sensor) calibration tasks, and vision-based simultaneous localization and mapping (SLAM). The state-of-the-art fiducial marker detection algorithms rely on the consistency of the ambient lighting. This paper introduces LiDARTag, a novel fiducial tag design and detection algorithm suitable for light detection and ranging (LiDAR) point clouds. The proposed method runs in real-time and can process data at 100 Hz, which is faster than the currently available LiDAR sensor frequencies. Because of the LiDAR sensors' nature, rapidly changing ambient lighting will not affect the detection of a LiDARTag; hence, the proposed fiducial marker can operate in a completely dark environment. In addition, the LiDARTag nicely complements and is compatible with existing visual fiducial markers, such as AprilTags, allowing for efficient multi-sensor fusion and calibration tasks. We further propose a concept of minimizing a fitting error between a point cloud and the marker's template to estimate the marker's pose. The proposed method achieves millimeter error in translation and a few degrees in rotation. Due to LiDAR returns' sparsity, the point cloud is lifted to a continuous function in a reproducing kernel Hilbert space where the inner product can be used to determine a marker's ID. The experimental results, verified by a motion capture system, confirm that the proposed method can reliably provide a tag's pose and unique ID code. The rejection of false positives is validated on the Google Cartographer dataset and the outdoor Honda H3D datasets. All implementations are coded in C++ and are available at: https://github.com/UMich-BipedLab/LiDARTag.

Video

Please checkout the introduction video. It highlights some important keypoints in the paper! <img src="./figure/LiDARTagVideoImg.png" width="960">

Quick View

LiDAR-based markers can be used in tandem with camera-based markers to address the issue of images being sensitive to ambient lighting. LiDARTags have been successfully applied to LiDAR-camera extrinsic calibration (paper, GitHub). This figure shows a visualization of LiDARTags of two different sizes in a full point cloud scan.

This system runs in real-time (over 100 Hz) while handling a full scan of the point cloud; it achieves millimeter accuracy in translation and a few degrees of error in rotation. The tag decoding accuracy is 99.7%.

<img src="./figure/first_fig1.png" width="640"> <img src="./figure/first_fig2.jpg" width="640">

Why LiDAR?

Robust to lighting!! The following shows LiDARTags are detected in several challenging lighting conditions:

Dingy environment
<img src="./figure/Dingy.gif" width="640">
Completely dark environment
<img src="./figure/CompletelyDark.gif" width="640">
Half tag being overexposed
<img src="./figure/HalfOverExposured.gif" width="640">
Rapid changing ambient light
<img src="./figure/RapidShort.gif" width="640">

Overall pipeline

The system contains three parts: tag detection, pose estimation, and tag decoding. The detection step takes an entire LiDAR scan (up to 120,000 points from a 32-Beam Velodyne ULTRA Puck LiDAR) and outputs collections of likely payload points of the LiDARTag. Next, a tag's optimal pose minimizes the <img src="https://render.githubusercontent.com/render/math?math=L_1">-inspired cost in (8), though the rotation of the tag about a normal vector to the tag may be off by <img src="https://render.githubusercontent.com/render/math?math=\pm 90^\circ"> or <img src="https://render.githubusercontent.com/render/math?math=180^\circ"> and will be resolved in the decoding process. The tag's ID is decoded with a pre-computed function library. The decoded tag removes the rotation ambiguity about the normal.

<img src="./figure/Pipelinev3.png" width="960">

Package Analysis

We present performance evaluations of the LiDARTag where ground truth data are provided by a motion capture system with 30 motion capture cameras. We also extensively analyze each step in the system with spacious outdoor and cluttered indoor environments. Additionally, we report the rate of false positives validated on the indoor Google Cartographer dataset and the outdoor Honda H3D datasets.

Pose and Decoding Analysis

Decoding accuracy of the RKHS method and pose accuracy of the fitting method. The ground truth is provided by a motion capture system with 30 motion capture cameras. The distance is in meters. The translation error is in millimeters and rotation error is the misalignment angle, (23), in degrees. <img src="./figure/Pose.png" width="960">

Computation Time of Each Step Analysis

This table averages all the datasets we collected and describes computation time of each step for indoors and outdoors.

<img src="./figure/PipelineSpeedAnalysis.png" width="960">

Cluster Rejection Analysis

This table takes into account all the data we collected and shows numbers of rejected clusters in each step in different scenes. Additionally, we also report false positive rejection for Google Cartographer dataset and Honda H3D datasets. <img src="./figure/ClusterRemoval.png" width="960">

Double-Sum Analysis

The original double sum in (18) is too slow to achieve a real-time application. This table compares different methods to compute the double sum, in which the TBB stands for Threading Building Blocks library from Intel. Additionally, we also apply a k-d tree data structure to speed up the querying process; the k-d tree, however, does not produce fast enough results. The unit in the table is milliseconds. <img src="./figure/DoubleSum.png" width="960">

False Positives Analysis

This table shows the numbers of false positive rejection of the proposed algorithm. We validated the rejection rate on the indoor Google Cartographer dataset and the outdoor Honda H3D datasets. The former has two VLP-16 Velodyne LiDAR and the latter has one 64-beam Velodyne LiDAR. <img src="./figure/FalsePositives.png" width="960">

Required Libraries / Packages

Those are the packages used in the LiDARTag package. It seems many but if you follow my steps, it should take you no more than 30 mins to instal them (including building time!). It took me awhile to get everything right. I summarize how I installed them here. However, you may choose any way you want to install them.

  1. Please install ROS Melodic.
  2. Please install TBB library. You may need to modify the CMakeLists.txt according to your installation.
  3. Please install NLopt. You may need to midify the CMakeLists.txt according to your installation.
  4. Please download LiDARTag_msgs and place them under your catkin workspace.
  5. The issue of Eigen version has been fixed by using an internal Eigen library instead of the system-wise Eigen library.
<!-- 5. Plesae ensure you have a correct Eigen Library on your system by downloading it from the [GitLab](https://gitlab.com/libeigen/eigen.git) and checkout `6f0f6f792e441c32727ed945686fefe02e6bdbc6`. Any commit older than this should also work. -->

Installation of Related Libraries

ROS Melodic

Please directly follow the instruction on the official website (here).

TBB library

Installation

Original TBB package from Intel does not support CMake; I, therefore, use another repository that supports CMake to make my life easier.

git clone https://github.com/wjakob/tbb
mkdir build;
cd build;
cmake ..;
cmake --build . --config Release -- -j 6;
sudo cmake --build . --target install
<!-- ###### Notes Ensure the followings in the CMakeList.txt are correct: 1. FindTBB.cmake is under _LiDARTag/cmake/_ 2. LIST(APPEND CMAKE_MODULE_PATH "YOUR_PATH/LiDARTag/cmake/") * Please change **YOUR_PATH** to your path to the LiDARTag package (something like this: catkin/src/LiDARTag/cmake). -->

NLopt library

Installation

Please direct follow the instruction on the official website (here) or as follow:

git clone git://github.com/stevengj/nlopt
cd nlopt
mkdir build
cd build
cmake ..
make
sudo make install

LiDARTag package

  1. Once you place LiDARTag_msgs under your catkin workspace and installed all the required libraries, you can directly catkin_make the package.
  2. source devel/setup.bash
  3. roslaunch lidartag LiDARTag_twotags.launch
  4. rosbag play -l -q bagfile.bag

Datasets and Results

Quantitative results:

If you would like to see how the tables in the paper are generated, please follow as below:

  1. Download this folder.
  2. Put them under LiDARTag/matlab/paper_data/
  3. Run genTable.m located at LiDARTag/matlab/

To regenerate results on the paper from scratch, please download the two datasets below:

  1. Please download bagfiles from here.
  2. Please download motion capture data from here
  3. change the output_path in the launch file
  4. roslaunch lidartag LiDARTag_threetags.launch
Note

The target sizes in the quantitative result folder are 1.22.

Qualitative results:

  1. Please download bagfiles from here.
  2. roslaunch lidartag LiDARTag_twotags.launch
Note

The target sizes in the qualitative result folder are 0.8051, 0.61.

False positive rejection:

Please download Google Cartographer dataset and Honda H3D datasets. We also provide different launch files (cartographer.launch, H3D.launch) for different datasets due to different published LiDAR topics and different output_path. I also wrote my own parsing script to pass bin files to rosbag. Please let me know if anyone needs it.

Running

  1. Please download qualitative bagfiles from here.
  2. catkin_make the package.
  3. source devel/setup.bash
  4. roslaunch lidartag LiDARTag_twotags.launch
  5. rosbag play -l -q bagfile.bag
  6. To see the results, rosrun rviz rviz. You can directly open LiDARTag.rviz under LiDARTag/rviz/ folder.
Notes

This package provides several launch files that you can directly run the package.
Please remember to change the tag_size_list in a launch file according to your target sizes or which bag file you are playing, or what marker sizes you have.
Different launch files:
-- LiDARTag_smallest.launch: only the smallest tag (0.61)
-- LiDARTag_twotags.launch: two smaller tags (0.61, 0.8)
-- LiDARTag_threetags.launch: all tags (0.8, 0.61, 1.22)
Please note that, the clearance around the markers should larger than <img src="https://render.githubusercontent.com/render/math?math=t\sqrt{2}/4">, where <img src="https://render.githubusercontent.com/render/math?math=t"> is the size of the largest marker. Therefore, it is recommended to use smaller tags in indoor environments.

Building Your Markers

<img src="./figure/LiDARTagSideView.png" height="280"> <img src="./figure/LiDARTagFrontView.png" height="280"> <img src="./figure/MarkerExample2.png" height="280">
We provide tag16h6c5 from AprilTag3 with three sizes (0.61, 0.85, 1.2).
If you want to use the provided markers, it is easy:

  1. Attach a fiducial marker to a squared cardboard or plexiglass and place the marker inside the yellow region.
    • Note: The sizes must be one of 0.61, 0.805, 1.22 meter, or you have to regenerate the function dictionary. If so, please follow here.
  2. Find a 3D object to support your marker. It could be a box or an easel.
    • Please note that, the clearance around the markers should larger than <img src="https://render.githubusercontent.com/render/math?math=t\sqrt{2}/4">, where <img src="https://render.githubusercontent.com/render/math?math=t"> is the size of the largest marker. Therefore, it is recommended to use smaller tags in indoor environments.
  3. Follow these steps to run the package.

Building Your Own Customized Markers

If you would like to use your own customized markers (i.e. different types of markers or different sizes), please follow these steps:
I. Build your function dictionary:

  1. git clone https://github.com/UMich-BipedLab/matlab_utils
  2. Add matlab_utils into build_LiDARTag_library.m or add matlab_utils into your MATLAB path.
  3. Edit opts.img_path in build_LiDARTag_library.m according to where you put images of your fiducial markers.
  4. Measure the size of your marker (<img src="https://render.githubusercontent.com/render/math?math=t">)
  5. Open build_LiDARTag_library.m in LiDARTag/matlab/function_dictionary/. Change opts.target_size_ to your marker size and run build_LiDARTag_library.m to generate your function library.
  6. Put the generated function dictuionary into LiDARTag/lib/
  7. When placing the generated function dictionary in LiDARTag/lib/, please put different sizes into different sub-folders (0, 1, 2, 3, ...) and put them in ascending order. For example, if you have three sizes (0.6, 0.8, 1.2), then you will have three sub-folders (0, 1, 2) inside the lib/ folder. Please place them as follow:
    • LiDARTag/lib/0/: put 0.6-size here
    • LiDARTag/lib/1/: put 0.8-size here
    • LiDARTag/lib/2/: put 1.2-size here

II. Follow Building Your Markers

Note

All the functions that are used for testing RKHS are all released in LiDARTag/matlab/function_dictionary/

Parameters of launch files

We split the parameters to two different launch files: LiDARTag_outdoor.launch and LiDARTag_master.launch. The front contains the most common tunables for different environments such as indoor or outdoor. The latter includes more parameters that you usually need to change for your system only once and just leave them there.

LiDARTag_outdoor.launch

feature clustering
cluster validation

LiDARTag_master.launch

System Mode

Debugging Mode

LiDAR Specification

Solvers for Pose Optimization

Decode Method

Tunable

feature clustering
cluster validation

Citations

The detail is described in: LiDARTag: A Real-Time Fiducial Tag for Point Clouds, Jiunn-Kai Huang, Shoutian Wang, Maani Ghaffari, and Jessy W. Grizzle. (PDF) (arXiv) (here)

@ARTICLE{HuangLiDARTag2020,
  author={Huang, Jiunn-Kai and Wang, Shoutian and Ghaffari, Maani and Grizzle, Jessy W.},
  journal={IEEE Robotics and Automation Letters}, 
  title={LiDARTag: A Real-Time Fiducial Tag System for Point Clouds},
  year={2021},
  volume={6},
  number={3},
  pages={4875-4882},
  doi={10.1109/LRA.2021.3070302}}