sorry I am not familiar with matlab

But for opencv I think you can calculate the homography matrix by the corner points coordinate in world and image. And then decompose the homography matrix

to obtain intrinsic and extrinsic matrix.

using the assumption that there is plane (tennis court) and that the tennish field is standardized (it has certain physical size), then you estimate the homography H by m>=4 points, and then you QR decompose the H for R and t for some arbitrary chosen camera matrix K (but it is assumed that you know the camera parameters, because you have the device), otherwise you can estimate the H, R, K from Beauguet "calibration toolbox for matlab" by m = 8 points where it is done by Marquardt-Levenberg non-linear optimization. Where for each assumption, that is projected from world to image until projection error is minimized.

Here is how I got parameters from three cameras views.

clear all; close all; clc;

images{1} = dir('Camera0/*.png');

images{2} = dir('Camera1/*.png');

images{3} = dir('Camera2/*.png');

load('GT.mat'); % ground truth

%individual color

humanCol = {[1 0 0],[0 1 0],[1 0 1]};

N_cams=length(actor2D{1}{1});

N_frames=length(actor2D{1});

for fr=704:1:N_frames %frame

for cam=1:1:N_cams

%load image data

im = imread([sprintf('Camera%d/',cam-1) images{cam}(fr).name]);

subplot(2,2,cam);

imshow(im); hold on;

for person=1:1:length(actor2D) %person (1,2,3)

xy= actor2D{person}{fr}{cam}; %2D joints

text(10,10,sprintf('Frame:%d, Camera:%d',fr,cam),'Color','r','FontSize',12);

for m=1:1:size(xy,1)

text(xy(m,1),xy(m,2), num2str(m), 'Color', humanCol{person});

if m==14

text(xy(m,1),xy(m,2)-10, sprintf('A%d',person), 'Color', humanCol{person}, 'FontSize', 12);

end

end

end

hold off;

end

pause(0.03);

end

I hope it can help

Try to use the Direct Linear Transformation (DLT) with well distributed points in different planes. Please check my profile for papers on DLT and a MATLAB code.

Regards,

Gamal