Been working for a while on this traffic camera detector.
Traffic Camera Detector - Scaled YOLOv4
Dataset —> ~2200+ annotated images
Right now it only works on pre-recorded images/video. Even if I would get it to work on the Nano(webcam) via ONNX the hassle would not be worth it because of the extreeeemly low FPS the Jetson Nano can produce on inference.
I know the dataset is sufficient because of the result I got via S-Yv4. I can get accurate detections on a couple of 100 meters.
Have also via TLT tested: Dashcam with extremly poor result. Likewise with TLT YOLOv4 and Darknet/19/53. Also Resnet10/18.
YOLOv5 has slighty better FPS but the accuracy is so bad that it would not work as an efficient detector.
Great learning experience but need alot more power than the Jetson Nano.