thatbrguy/Pedestrian-Detection ? reverse-engineered prompt

Build me a pedestrian detection project that can train on the TownCentre dataset and then run detections on video frames or images. I want it to use the TensorFlow Object Detection API, with scripts that can pull the dataset, extract frames from the source video, turn the ground truth into XML annotations, and create the training records automatically.

It should also include the option to use Nanonets for training and predictions, since this repo compares those workflows too. Make the inference side able to run across multiple GPUs in parallel so I can see better speed and frame rate when processing a video. If it makes sense, include a simple way to show the detected people on output frames and save results so I can compare performance.

Please keep it practical and easy to run, with clear setup steps and sensible defaults. If you need to check any current docs or commands online, go ahead and do that.