Horse Finish Line Detection

I need a YOLO-based vision pipeline that can spot the exact moment a horse cart’s nose or leading leg breaks a fixed virtual line at the finish and then instantly assign race positions. Every cart stays in its own lane, so lane-switching logic is unnecessary; the job is all about reliable object detection, crisp line-cross logic, and live ranking.

Model & detection
I’m flexible on the YOLO version—v3, v4, v5, or whichever you feel delivers the best speed-to-accuracy mix on HD footage & it needs to be used for youtube live. What matters is that the trained model consistently labels two classes per horse: Nose and Leg. Those detections feed a simple, lane-aware crossing algorithm that timestamps the first pixel intersection with the finish-line ROI.

Finish-line setup
The detection zone will be a static, calibrated strip across the track. Once positioned, it remains fixed for the whole meet, so you only need one homography or coordinate transform for alignment.

What I expect you to hand over
• Training notebook or script with clear instructions to retrain on new footage
• Inference script (Python preferred, OpenCV OK) that ingests live or recorded video, draws detections, logs crossing times, and prints/exports a ranked list in real time
• Configuration files, class labels, and weights
• A short README walking me through installation, calibration of the finish-line ROI, and how to start a race-day session

Acceptance
I’ll run your code on sample race footage: if the printed ranking matches the actual finish order frame-by-frame and latency stays under 200 ms per frame on a consumer-grade GPU, we’re good.