Cal_PnP

This repository contains our C++ implementation of semi-automatic camera calibration based on Perspective-n-Point (PnP). It can be used to compute the homography matrix and manually correct radial distortion.

Introduction

This package is designed for computing camera homography matrix. The input includes a set of 3D points on the 3D ground plane and the corresponding 2D pixel points. We adopt the Perspective-n-Point (PnP) for camera calibration. The user can choose from different approaches such as (1) a regular method using all the points, (2) RANSAC-based robust method or (3) Least-Median robust method. The package also includes tools for manual 2D points selection.

Coding Structure

./obj/ folder: Libraries of .cpp files
./src/ folder: Source code
./data/ folder: Example input files
1. cfg.json: Configuration parameters in JSON format
2. frm.jpg: Input frame image

How to Build

Download and make the OpenCV library. A tutorial is given here.
Compile using g++ in Linux environment. If you are new to g++ compilation with OpenCV, please refer to this link. In the command window, you can cd to the directory of ./Cal_PnP/ and use the following command to compile our source code, where bin is the executable file generated. Note that you may need to add sudo at the beginning to grant the admin permission.
```
g++ -I/usr/local/include/ -L/usr/local/lib/ -g -o bin ./src/main.cpp ./src/Cfg.cpp ./src/CamCal.cpp -lopencv_core -lopencv_imgproc -lopencv_highgui -lopencv_imgcodecs -lopencv_calib3d -lm
```

How to Use

Set the corresponding input/output paths in the configuration file if necessary.
Provide the coordinates of at least 4 3D points on the ground plane (or any plane of interest) at cal3dPtLs. One way is to find the GPS coordinates using Google Maps.

If the user wants to select the corresponding 2D pixel points on the image, set calSel2dPtFlg to 1. A new window called selector of 2D points pops out. The user can click on the image to select each 2D point. A blue circle stands for each click. After the selection is done, click o. Please make sure the number of selected 2D points should be the same as the provided 3D points (in the same order). During the selection, if mis-clicking on wrong places, the user can press r. All the markers will be cleared, and s/he can start over.

If there exists redial distortion, the frame image can be manually corrected by by setting calDistFlg to 1 and providing the distortion coefficients (calDistCoeff) and intrinsic camera parameters (calFocLen and calPrinPt).
The user can choose to use different methods for camera calibration by setting calTyp, whose introduction is at this link. Note that when using RANSAC-based robust method, the threshold parameter needs to be provided at calRansacReprojThld. When setting calTyp as -1, all the methods will be conducted and evaluated, the one with the minimum reprojection error will be chosen.
The output text file shows the 3x3 homography matrix at the first line. If the correction of radial distortion is conducted, the 3x3 intrinsic parameter matrix and 1x4 distortion coefficients are also printed. Finally, the reprojection error in pixels is printed as well.
The user can choose to output the display image (with a colorful grid on the 3D plane) by setting outCalDispFlg to 1. The blue circles show the input 2D pixel points. The red circles show the corresponding points back projected from 3D. The distance between their centers indicates the reprojection error.

Run the executable file.
```
./bin
```

References

The code was applied to the generation of baseline camera calibration results for the CityFlow benchmark used in the AI City Challenge Workshops. Please consider to cite these papers in your publications if it helps your research:

@inproceedings{Tang18AIC,
  author = {Zheng Tang and Gaoang Wang and Hao Xiao and Aotian Zheng and Jenq-Neng Hwang},
  title = {Single-camera and inter-camera vehicle tracking and {3D} speed estimation based on fusion of visual and semantic features},
  booktitle = {Proc. CVPR Workshops},
  pages = {108--115}, 
  year = {2018}
}

@misc{Tang17AIC,
  author = {Zheng Tang and Gaoang Wang and Tao Liu and Young-Gun Lee and Adwin Jahn and Xu Liu and Xiaodong He and Jenq-Neng Hwang},
  title = {Multiple-kernel based vehicle tracking using {3D} deformable model and camera self-calibration},
  howpublished = {arXiv:1708.06831},
  year = {2017}
}

Disclaimer

For any question you can contact Zheng (Thomas) Tang.

enteguo / Cal_PnP