That certainly is some terrible green screen footage.
Personally I would first mask the windows exactly, as if the people were not there at all. If there is any camera movement or jitter, track it and apply it to that mask. (This way you have less concern with accidentally including the green spill into the key.)
Then I would try and key just the people moving in front of the windows, one window at a time.
In other words, I would build the mask from four layers, the clean windows, masked by hand, then left, back and right (people in front of the) windows separately with some rather extreme keying to get just the silhouettes of the people, wouldn’t care that much if the key eats some into the people (using choke). The result of this would be a black and white mask, which is then used as a track matte for the original footage.
Change color or perhaps Selective color for the green spill on the original footage. Or even an HLS selector in Color Finesse if that doesn’t do it.
If you want some detail back in the windows, fake it and put whole new layers there to replace the glasses.