Today we put everything together and gave it a try. 

Its so much fun to use this thing and everything works quiet well.

We have just a few issues.
- As the cameras of course can not cross-eye (dont know how to say that in english), you will see very close object double.

- We have some problems with the performance. With Object Recognition turned on, we have ~ 70-80% CPU load on a quad-core machine. I had to rearrange some thread priorities to stop the GUI from freezing. Its working but I want to optimize it. I maybe need to rethink how to process the images, maybe I also will try another framework for the cameras.