When any moving platform flies over a piece of ground, capturing sequential (perhaps overlapping) imagery, the image perspective changes. The distortion caused by this is removed by resampling each image to a common coordinate system, usually defined by previously known coordinates of distinct control features visible on the images. This process, called rectification, is generally done with an affine transform where each x,y on each image is transformed to common coordinates X,Y by an equation similar to

X = ax + by + c and Y = dx +ey + f

The constants a through e are determined by solving simultaneously the linear equations

X(i) = ax(i) + by(i) + c for all i control points

(the same is done for the Ys to get d,e,f) .

where the X(i),Y(i) and x(i),y(i) are the known ground coordinates and pixel values, respectively, of the control points. Because of inaccuracies in the original data, the system is overconstrained so you have to do a least-squares solution which yields an RMS error for the transformed values. You need more equations than there are unknowns, the more the better, and the solution is always approximate (which is why the images wiggle when they are flashed in sequence). They never overlay perfectly.

If the accuracy of your control and precision of the optical system justifies it, higher order (non linear) transforms and map projections can be used to reduce this image shift, but it never goes away altogether.

I hope that helps.