ifzhang/FairMOT — reverse-engineered prompt

Please build me a FairMOT style Python project for tracking people in videos. I want to point it at a video file or a MOT Challenge image sequence and get back an annotated video plus tracking results, with each person boxed and given a consistent ID as they move around.

Use the pretrained FairMOT baseline if possible, and include clear setup steps for Python 3.8, PyTorch, the DCNv2 backbone piece, model files, and ffmpeg so I can run the demo on my machine. Make the commands simple, like run a demo on videos in a videos folder, evaluate on MOT data, and optionally prepare labels if I download the datasets later.

Keep the code organized, add sensible defaults, and include a README that explains where to put models and datasets. If anything has changed since the original FairMOT instructions, look up the current docs online and make it work.