nisanimsi/ugv-real-time-object-detection ? reverse-engineered prompt

Build me a simple computer vision project for a UGV that can detect three things in outdoor camera footage, person, vehicle, and obstacle.

I want the full workflow in one place, from taking frames out of video or RTSP camera streams, to annotating data in CVAT, exporting it in a YOLO friendly format, training a model, and then testing it on images, videos, webcam input, and live RTSP streams. Please include a clear dataset setup, training and evaluation scripts or commands, and a way to measure things like precision, recall, mAP, FPS, and latency.

It should also make it easy to review failure cases like small objects, low light, blur, or occlusion, so I can improve the dataset over time. Use Python, OpenCV, YOLO, PyTorch, and Ultralytics if that fits best. If you need anything current from the docs, look it up online first.