---

LibQPEP

A library that deals with generalized quadratic pose estimation problems (QPEPs). The algorithm aims to obtain globally optimal pose together with globally optimal covariance estimates. Typically it can solve the following problems:

1. Perspective-n-Points (PnP)
2. Perspective-n-Lines (PnL)
3. Perspective-n-Points and Lines (PnPL)
4. Hand-eye Calibration
5. Point-to-plane Registration
6. Conics-based Camera Pose Estimation
7. Multi-robot relative pose problem from range measurements
8. Forward kinematics of parallel robots
9. Multi-GNSS attitude determination problem

Authors

Affiliation: Hong Kong University of Science and Technology (HKUST)

Contributor: Jin Wu (https://github.com/zarathustr), [email protected]

Co-Contributor: Xiangcheng Hu, HKUST (https://github.com/JokerJohn)

Usage

The C++ codes are built using CMake toolkit under the C++11 programming standard. The codes have been verified on the Ubuntu 14.04/16.04/18.04 (GCC Compilers 5.0 ~ 10.0), Mac OS X 10.5.8/10.6.8/10.7.5/10.8.5/10.9.5/10.10/10.12/10.14/10.15 (Clang Compilers 3 ~ 11). We support multiple architectures, including x86, armhf, arm64, etc.

In the C++ code, the file main.cpp contains demos of pose and covariance estimation. The function QPEP_grobner solves the QPEP via Groebner-basis elimination by Larsson et al. https://github.com/vlarsson. Using QPEP_lm_single, the solved pose will be refined by the Levenberg-Marquadt (LM) iteration. Finally, the function csdp_cov estimates the covariance information.

The LibQPEP can be accelerated by many parallel-computing approaches, including BLAS, LAPACK, CUBLAS (CUDA), Metal Performance Shaders, and OpenCL.

The LibQPEP now supports multi-architecture hardwares including Nvidia TK1/TX1/TX2/Xavier, RK3399, NXP iMX.6x Series, and softwares including MATLAB R2007~R2022.

Dependencies

1. Mandatory dependencies are: X11, LAPACK, BLAS, Eigen3. For OSX and macOS users, please feel free to install the dependencies via Homebrew or MacPorts.
2. OpenCV is optional. However, if you need visualization of covariances, OpenCV must be installed. We support OpenCV 2.x to 4.x.
3. To enable OpenCL support, please install ViennaCL (https://github.com/viennacl/viennacl-dev). Also note that, the OpenCL is efficient only if the graphics devices are based on either Intel/ARM Graphics or Intel Compute Stick. For NVidia/AMD/ATI graphics card, the memory copy processing time is significantly longer.

C++ Compilation

git clone https://github.com/zarathustr/LibQPEP
mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make install

Note: Using Clang compiler on Unbuntu 16.04, 18.04 will be much faster than GCC! For Ubuntu 20.04 and macOS, the execution speed will be consistent. To use the Clang compiler in CMake, follow these instructions:

CC=/usr/bin/clang CXX=/usr/bin/clang cmake ..
make install

where /usr/bin/clang directs to the path of the installed Clang compiler.

Demo Program

Just run

./LibQPEP-test

To specify an algorithm to test, please use:

./LibQPEP-test [PnP | PtoP | Hand-eye]

A successful execution of ./LibQPEP-test PtoP (Point-to-Plane Registration) will output:

To include a certain data file for test, please run:

./LibQPEP-test [PnP | PtoP | Hand-eye] data_file

For instance, if you would like to try dataset pTop_data-2500pt-1.txt in data folder, use the following command:

./LibQPEP-test PtoP /data/pTop_data-2500pt-1.txt

which will output:

Library Installation

This depends on the CMake prefix you set by -DCMAKE_INSTALL_PREFIX=path_to_your_install. The standard path is /usr/local. Use make install to install the headers and libraries along with CMake and pkgconfig files.

Troubleshooting

1. On the Linux machines, if you encounter the following error:

No rule to make target '/usr/local/share/LibQPEP/cmake/../../../lib/libLibQPEP.a'

Please use the following command:

sudo ln -s /usr/local/share/LibQPEP/cmake/../../../lib/libLibQPEP.so /usr/local/share/LibQPEP/cmake/../../../lib/libLibQPEP.a

provided that your CMake install prefix is /usr/local.

2. To specify a certain version of OpenCV, please follow:

   cmake .. -DOPENCV_VER=X.Y.Z -DOPENCV_PATH=your/path/to/opencv
   

   where OPENCV_VER is the exact version of your OpenCV (e.g. 3.2.0, 4.5.0, 2.4.9), while OPENCV_PATH directs to the install prefix path, e.g. /usr/local for default.

3. For OpenCV 3.2.0 installed by Debian APT package manager, the intrinsic usage of fast math will be a little bit inaccurate, which induces a bit inconsistency of covariance ellipses with statistical and QPEP tests.

MATLAB Demo Kit

The MATLAB of version over R2007b is required for proper evaluation. The MATLAB demo kit mainly consists of examples showing how QPEPs are constructed and solved. Three files syms_hand_eye.m, syms_pnp.m, syms_pTop.m contain symbolic generators for expression functions of hand-eye calibration, PnP and point-to-plane registration problems. Two test files test_rel_att.m and test_stewart.m illustrate how range-based relative pose problem and the forward kinematics problem of Stewart platform can be transformed into QPEPs. The final three files test_cov_hand_eye.m, test_cov_pnp.m and test_cov_pTop.m consist of globally optimal solutions and covariance estimation. The comparison with method of Nguyen et al. is shown in nguyen_covariance.m. This comparison with Nguyen et al. requires installation of Python 2.7, with packages of numpy, trimesh, scipy, and matplotlib (2.2.4, strict!). The implementation is ported directly from the original authors' repository https://github.com/dinhhuy2109/python-cope. Users can change the path in the script to choose the proper Python interpreter. In covariance estimation codes of MATLAB, we use the SeDuMi as a general optimizer, since it is free and flexible.

Applications (C++ & ROS Codes)

1. Point-to-plane Registration: https://github.com/zarathustr/libpointmatcher_QPEP (Can be used in ROS with https://github.com/ethz-asl/ethzasl_icp_mapping and https://github.com/ANYbotics/pointmatcher-ros)
2. Perspective-n-Points for SLAM Cloop Closure: DSO: https://github.com/zarathustr/DSO-QPEP-Loop-Closure VINS-Mono: https://github.com/zarathustr/VINS-Mono-QPEP
3. Hand-eye Calibration (ROS): https://github.com/zarathustr/Global-Hand-Eye-Calib-Toolkit
4. Event-Camera/Lidar extrainsic calibration: https://github.com/HKUSTGZ-IADC/LCECalib

Extensions

1. For single-precision (float) and long-double implementation, please go to https://github.com/zarathustr/QPEP-MultiPrecision.
2. For solving QPEPs with random sample consensus (RANSAC), please go to https://github.com/zarathustr/QPEP-RANSAC.
3. For arbitrary-precision implementation, please go to https://github.com/zarathustr/QPEP-MPFR.

Supplementary Material

The usage of the theoretical proofs and mapping toolbox can be found in the supplementary file suppl.pdf

How to Contribute to LibQPEP?

The C++ version of LibQPEP originates from its MATLAB version codes in the code folder. To define a quadratic pose estimation problem (QPEP), we recommend that the problem can be written in the form of scalar objective function L such that the pose T on SE(3) is optimized via argmin J, subject to the SO(3) constraints in T. Then, please follow the syms* files in the root of MATLAB demo codes to generate the required numerical matrices for QPEP. After that, you may need to convert the matrices into C++ callable functions. Please refer to misc_*_funcs.cpp files for such a conversion. Finally, you may solve the problem via the provided solvers. If one needs the covariance estimation, please follow the steps in main.cpp, where an example of the covariance estimation of point-to-plane registration has been presented. We expect that more problems can be solved in the QPEP fashion!

References

1. Wu, J., et al. (2022) Quadratic Pose Estimation Problems: Globally Optimal Solutions, Solvability/Observability Analysis and Uncertainty Description, IEEE Transactions on Robotics, https://doi.org/10.1109/TRO.2022.3155880.

2. Wu, J., et al. (2020). Globally Optimal Symbolic Hand-Eye Calibration. IEEE/ASME Transactions on Mechatronics, https://doi.org/10.1109/TMECH.2020.3019306