I didn't expect you to know what it means. I suggested that you find out what it means. Terms get bandied about so that you can go and find out what they mean.
Of course... but not as productive as when you follow up with searching, finding and reading some existing literature.
I'm glad you discovered that auto-correlation using X-OR doesn't work very well, but any intuition would have told you the same thing.
To handle image distortion, camera motion, seeing the same thing from slightly different angles, you need to do feature extraction and object recognition. A typical approach is to differentiate the image in two or more dimensions, pick the N highest peaks (or even better pick where a peak in one direction is a trough in the orthogonal direction, aka a corner) and then for each such point, make it into a triplet with its closest two neighbours. Figure out the angle and two side lengths of the triangle that forms, and index each triplet using those as lookup keys in an approximate lookup table. Now use the most prominent features to search in the same table made from the previous image, and you have a collection of probable match triplets. For each probably match, compute the transformation that moves the old triplet to overlay the new triplet. Out of all that, you'll get a massive cluster of similar transformations, and a lot of outliers. Discard all the outliers, and use a least-squares fit over the transformations in the cluster. That gives you a transformation from the perspective of the previous image to the current one with very high confidence. You can even use the systematic divergence in the errors to calculate what type of lens distortion is present, as Autostitch does.
But you could have learned all that for yourself the same way I did, by starting with Google.
Clifford Heath.