Issues

It doesn't really care what pt file is included and it always use YOLO's model id, for example if id 1 is apple, it still say person. maybe extract class list from yolo's .pt somehow?
It read frame a bit too fast. it say it's infering at 20-ish fps but the actual camera is only 5 fps or so
Potential race condition issue when multiple camera try to init with the same unconverted model.
Buffer Size for EACH CAMERA should be one, batch is for when processing multiple cameras, when new frame comes in, replace the old one at the old index if exist. This way the real time requirement is satisfied. We need a data structure to track this in addition to ring buffer tho.
Buffer should flush after TARGET_FRAME_INTERVAL_MS
Blurry asyncio archtecture, require documentations