How to fine tune a 2D projective transformation

I'm wondering if there are any codes out there that use an automated process to fine tune a projective image transformation -- similar to how imregister uses an optimizer and metric to estimate an affine transformation.

I am trying to register two thermal images taken a second apart from one another during a drone flight. Using control points, the command cpcorr, and the command fitgeotrans, I can register the second on top of the first with a correlation coefficient of 0.86, but I'm convinced the images can fit better than this.

I've tried to put together a script to optimize the image correlation, by treating each term in the transformation matrix as an independent variable and using fmincon -- but so far I haven't had success with that route.