This question is discussed quite a lot here. You can have multiple cameras connected to the board as long as they fit in the CSI lanes available on the SoM. In my project we have 3.
Then provided you write the appropriate V4L2 driver for the camera (this is mainly getting configuration register tables from somewhere) you can use a program of your choice to get raw frames from the camera, yavta being a good choice. This means that if your camera is UYVY one, you get the frame in this exact format, or Bayer sensor would give BGBG/GRGR sequence in the frame.
However, in order to use all features that gstreamer or Vision Works provides you need some close source components from Nvidia that transform YUV or Bayer to a an acceptable format (for example NV12). These components I found have showstopper compatibility bugs that still make cameras other than the default ov5693 (the evaluation kit is shipped w/ it) unusable.