Hi, I am not sure, if this is still relevant for you, but I can still share some experience here, if someone else was interested, too.
The only way we managed to save some video sequences to the target HDD was to reshape the video signal to a row vector and pass the vector to a File Scope. (Reshaping to a column vector caused that the File Scope was only saving the first element of the vector.)
I have to emphasize that this approach is only feasible for low-resolution videos. The higher resolution the video has, the longer the build and initialization of the model on the target takes. For a video with resolution 256x256 pixels the whole process from building to launching the target application took several minutes on an i5-powered machine.
I am quite curious whether MathWorks improved the video support in the latest Simulink Real-Time, which replaced xPC Target.