Camera calibration

Notes: Please install GitHub with MathJax to view LaTex formulation, Thank you.

• Theory
• Camera calibration with SIFT detection and match algorithm

Theory

• Definition: A camera projects 3D world-points onto the 2D image plane.

• Calibration: Finding the quantities internal to the camera that affect this imaging process

  • Image center
  • Focal length
  • Lens distortion parameters

• Basic coordinate

  • world coordinate
  • camera coordinate
  • image coordinate
  • plane coordinate (defined by myself for easily explaining)

• Main procedures

  • stage 1: from world coordinate to camera coordinate, this is from 3D points to 3D points including $R$ and $t$ etc. parameters.
  • stage 2: from camera coordinate to image coordinate, this is from 3D to 2D points including $K$ (camera intrinsic) etc. parameters.

• From world coordinate to camera coordinate

  • Coordinate introduction

  

  From world coordinate to camera coordinate needs two operations: translate and rotation. It can be simply described: firstly, translate the origin of word to the camera coordinate location and then rotate it for aligning the coordinates. $$ \widetilde{X}_{cam} = R(\widetilde{X}-\widetilde{C}) $$ Where,

  ​ $R$ rotation matrix

  $\widetilde{X}$ and $\widetilde{C}$ is the location in the world coordinate respectively

  ​ $\widetilde{X}_{cam}$ is the location in the camera coordinate

  • From world coordinate to camera coordinate

  

  We set $R=R_x(\psi)R_y(\varphi)R_z(\theta)$ therefore, $$ X_{cam}=\left( \begin{array}{ccc}X_{cam}\\1\end{array}\right)=\left[\begin{array}{ccc}R&−R\widetilde{C}\\0&1\end{array}\right]\left( \begin{array}{ccc}\widetilde{X}\\1\end{array}\right)=\left[\begin{array}{ccc}R&−R\widetilde{C}\\0&1\end{array}\right]X $$ Where we also can say $t=R\widetilde{C}$ and the $X$ we will discuss later.

• From camera coordinate to image coordinate

  • Coordinate introduction

  

  Where,

  $C$ is camera center

  $Z$ is principal axis

  The plane of $p$ is image plane, it is a 2D plane

  $p$ is principal point

  $f$ is focal length, the distance between $C$ and $p$

  The plane coordinate $xpy$ is paralleled with the camera coordinate $XCY$ and the measure value is $metre(m)$

  The image coordinate usually said the relative coordinate to plane coordinate, its origins is in the left-bottom corner, measured by the numbers of $pixel$

  • From camera coordinate to plane coordinate

  Given $X$ in the our real world, the corresponding point in the camera coordinate is $x$. Requirement from $X(X,Y,Z)$ to $x(x,y)$. Note: $(X,Y,Z)$ is in the camera coordinate and $(x,y)$ is in the plane coordinate (different with image coordinate, the origins is different). We can get: $$ x=f\cdot\frac{X}{Z} \ y=f\cdot\frac{Y}{Z} \ (X,Y,Z)↦(f\cdot\frac{X}{Z},f\cdot\frac{Y}{Z}) $$ In matrix, $$ \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right)↦ \left(\begin{array}{ccc}f\cdot{X}\\f\cdot{Y}\\Z\end{array}\right)= \left[\begin{array}{ccc}f & 0 & 0 & 0 \\ 0 & f & 0 & 0\\0&0&1&0\end{array}\right] \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right) $$ Abbreviation, $P$ is the projection matrix but not complete now. $$ x=PX $$

  • Adding offset

  We have converted object point from camera coordinate to plane coordinate, but our target is get the image coordinate.

  

  We know, $p$ is the origin of plane coordinate but it is $(p_x,p_y)$ in the image coordinate. $$ (X,Y,Z)↦(f\cdot\frac{X}{Z}+p_x,f\cdot\frac{Y}{Z}+p_y) $$ Matrix, (can divided by $Z$) $$ \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right)↦ \left(\begin{array}{ccc}f\cdot{X}+Zp_x\\f\cdot{Y}+Zp_y\\Z\end{array}\right)= \left[\begin{array}{ccc}f & 0 & p_x & 0 \\ 0 & f & p_y & 0\\0&0&1&0\end{array}\right] \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right) $$ Therefore, $$ \left(\begin{array}{ccc}fX+Zp_x\\fY+Zp_y\\Z\end{array}\right)= \left[\begin{array}{ccc}f & 0 & p_x \\ 0 & f & p_y \\0&0&1\end{array}\right] \left[\begin{array}{ccc}1 & 0 & 0&0 \\ 0 & 1 & 0 &0\\0&0&1&0\end{array}\right] \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right) $$ Here, $K$ named intrinsic parameters is defined by $$ K=\left[\begin{array}{ccc}f & 0 & p_x \\ 0 & f & p_y \\0&0&1\end{array}\right] $$ And $P$, we can assume the rotation matrix $R$ is $I$, and the translate matrix $t$ is $0$ $$ P=K*[I|0] $$

  • From plane coordinate to image coordinate

  $$ x=K[I|0]X_{cam}=K[R|-R\widetilde{C}]X $$

  The final $p$ is $$ P=K[R|t] $$ Where $R, t$ is named the camera extrinsic parameters.

  • Convert the image coordinate by pixel

  In the image coordinate, it usually use the pixel to measure not m. So, we need another step. We set

  $m_{x}$ is the numbers of pixel of $1m$ in the horizontal and,

  $m_{y}$ denotes the numbers of pixels of $1m$ in the

  Thus, we can get the area of one pixel ($m^2$), $$ \frac{1}{m_{x}}\times\frac{1}{m_y} $$ The pixel can be not the square, it also can be rectangle. We update $K$ $$ K=\left[\begin{array}{ccc}m_{x} & 0 & 0 \\ 0 & m_{y} & 0 \\0&0&1\end{array}\right] \left[\begin{array}{ccc}f & 0 & p_x \\ 0 & f & p_y \\0&0&1\end{array}\right]= \left[\begin{array}{ccc}f_{x} & 0 & c_x \\ 0 & f_{y} & c_y \\0&0&1\end{array}\right] $$ Generally, when we use camera intrinsic parameters $K$ to calculate, the provided parameters is converted values, e.g. it will provide $f_{x},f_{y},c_{x},c_{y}$.

  • Summary

  The final formulation is, $$ \left(\begin{array}{ccc}xZ\\yZ\\Z\end{array}\right)= \left[\begin{array}{ccc}f_{x} & 0 & c_x \\ 0 & f_{y} & c_y \\0&0&1\end{array}\right] [R_{3x3}|t_{3x1}] \left(\begin{array}{ccc}X\\Y\\Z\\1\end{array}\right) $$ Thus, we can get $x(x,y)$ $$ \left(\begin{array}{ccc}xZ\\yZ\\Z\end{array}\right)\frac{1}{Z}↦ \left(\begin{array}{ccc}x\\y\\1\end{array}\right) $$

• Considering Distortion Parameters

  The distortion Parameters includes Radial distortion and tangential distortion

  • Radial distortion includes barrel distortion, pincushion distortion and mustache distortion.

    

    It can be corrected by $$ x_{corr}=x_{dis}(1+k_{1}r^2+k_{2}r^4+k_{3}r^6)\ y_{corr}=y_{dis}(1+k_{1}r^2+k_{2}r^4+k_{3}r^6) $$

  • tangential distortion

  

  It can be corrected by $$ x_{corr}=x_{dis}+[2p_{1}xy+p_{2}(r^2+2x^2)]\ y_{corr}=y_{dis}+[p_{1}(r^2+2y^2)+2p_{2}xy] $$ Where,

  $x_{dis},y_{dis}$ represent the distorted coordinate

  $x_{corr},y_{corr}$ represent the corrected coordinate

  $k_{1},k_{2},k_{3}$ is the radial distortion parameters

  $p_{1},p_{2}$ is the tangential distortion parameters

  Therefore, we get 5 distortion parameters $$ D=(k_{1},k_{2},p_{1},p_{2},k_{3}) $$

Camera calibration with SIFT detection and match algorithm

• Using opencv api to calibrate camera, you need to change img_root in calibration_opencv.py

  python calibration_opencv.py
  

• Maybe you want to use harris detector to try using your image.

  python harris_detector.py
  

• Self-implement main steps: (see scripts for more information and feel free to modify)

  • Use your camera or phone take several pictures (at least 4), put in the chessboards_imgs directory

  • Run detect_keypoints_sift.py for each pattern images in chessboards_imgs directory to get related keypoints and descriptors. For example:

    python detect_keypoints_sift.py --in_fname ./chessboards_imgs/left01.jpg --out_fname keypoints_dir/left01.txt
    

  • Run sift_match.py for each pattern images in chessboards_imgs directory to get related keypoints and descriptors. For example:

     python match_keypoints.py --in_fname_1 ./chessboards_imgs/left01.jpg --in_fname_2 ./chessboards_imgs/left01.jpg --out_fname match_left01_to_left02.txt
    

  • Run calibrate_intrinsics.py to calculate camera intrinsics.

    python calibrate_intrinsics.py --pattern_key_fname keypoints_dir/left01.txt --img_keypoints_root ./keypoints_dir/ --select_match_keypoints_root ./select_match_keypoints_dir/ --example_image ./meilan_note6_resized/note01.jpg
    

• Do perspective_transform using opencv api, you need to change the image path in the perspective_transform.py

  python perspective_transform.py
  

  Notes: need to adjust related parameters to get a better result.

[TODO] Implement the function calibrateCamera.

If you have any question or find any problems, please contact me or issue.

References

• https://blog.csdn.net/Peng___Peng/article/details/51725219
• https://blog.csdn.net/honyniu/article/details/51004397
• https://blog.csdn.net/mightbxg/article/details/79363699
• https://github.com/ksimek/sift_calibration
• https://github.com/rmislam/PythonSIFT/blob/master/siftdetector.py