Is it possible to train an image / dataset of a image (captured by the camera) on an actual device? Or is this just possible doing with the command line tools?
Is there an API existing for this or do i have to compile the genTexData tool for the devices?
ARToolkit does not support auto-training of markers. In fact the command line tool can take a long time for complex or large markers.
It is possible, but you would need to do it yourself. As you suggested, compiling genTexData for the device is the way to go (or putting it into a server and accessing it via http).