you’ll probably need a calibration pattern that is spatial, i.e. not flat. and it, or parts of it, need to be viewable by all cameras at once. that means you need to be able to detect parts of it.
I think people use aruco for these purposes.
you’ll probably need to calibrate the pattern itself because it’s hard to impossible to manufacture rigid enough patterns of that size.
you’ll calibrate one camera on its own to remove its distortion. then you use that camera to measure the pattern’s geometry, and that’ll hopefully agree with manual measurements (yard stick/tape measure, especially diagonally/through space).
once you have an accurate model of your pattern, you can show it to your cameras, and detect parts of the pattern in all cameras, and use those detected points together with the model points.
do literature research on how those 360 degree multi-cam “soccer balls” are calibrated… or how the composite surround view of autonomous vehicles is calibrated.