vision4robotics/MobileMVT — reverse-engineered prompt

Build me a Python project that can track multiple vessels in drone or UAV video in real time. I want to be able to point it at a video plus a camera parameter file, run one command, and get back an annotated output video with stable boat IDs and trajectories. It should be aimed at maritime surveillance, work well when boats appear at different sizes and viewing angles, and handle camera motion from the drone so the tracking does not jump around too much.

Please include a ready to run demo using the sample video and sample camera parameters, and make sure the output is saved automatically. I also want the repo set up so I can train the detector on the VesselMOT dataset later if I have the data, with a clear training command and dataset format. Keep the structure clean, installable from requirements, and add simple instructions for how to run demo and training. If anything is unclear, check the current docs online and fill in the sensible defaults.