I came up with an idea to precalculate the slower parts of the silouette detection, which basicly works the same way as before but A LOT faster. It is quite memory consuming, but I doubt that this will ever be a problem. That precalculation is kinda slow, but only has to be done once and is independant of light and model positions and orientations.
The spacecraft on the following screenshot was running with about 0.1fps with shadows before, so I´d say that this is a big improvement (the scene should have about 2.5k triangles, maybe a bit more):


And another one (silouette detection is really fast, but rendering of the volume needs some speed up):


And one more