[zipy124]

Not an expert but I imagine the problem with using vision is the problem of angular error propagation. That is the angle-> distance error problem:

linear error≈Rtan(Δθ)≈RΔθ

Here linear error is the error in position, R is the distance from the observer to the target and θ is the angle error. You would need incredibly good optics and resolution to minimise angular error and thus linear error.

[_diyar]

Yes but the errors would be uncorrelated and if you could use many different measurements you should be able to estimate with reasonable certainty.

[zipy124]

Given the large R (Even starlink satellites are far away compared to positional accuracy needed) and probably pretty large angular error, your still talking error of kilometres, though with good optics you may be able to get this down to hundreds of meters. The other problem is knowing the position isn't trivial, since their position drifts relative to predicted and isn't known very precisely.

There is also the problem that as R shrinks, speed increases relative to you.