Ok, you can’t directly connect a camera to or run computer vision code on the autopilot board, so that will need to be processed on the companion computer, which can then guide the vehicle on where to move by sending relevant MAVLink messages (using Pymavlink or some other language implementation) 
Depending on what you want from this, it may be possible to visually follow something back to the base station. If there’s nothing to follow then it will need some kind of positioning reference/sensor, which could be self-contained on the vehicle (e.g. a DVL), or could use one or more external acoustic beacons that are detected by receivers on the vehicle (possibly a direct acoustic positioning system, but that’s not essential, especially if you only have one location you want/need to get to).