Very interesting...
Must be possible by automatically selecting pixels that are far from their average value (I mean, if you take 10 pictures, each pixel has an average value over these 10 pictures, and if in one image a pixel has a value which is far from its average value, then you keep it). To prevent noise you can also do it only if the connected component of such pixels is large enough (or just the largest is there is only one subject), and to prevent strange superimpositions only add independant connected components.
Does such software already exist?
