Vehicle counting for Little Cottonwood Canyon (SR-210, Utah) from public UDOT traffic cameras, using RF-DETR running locally via Roboflow Inference.

Every 10 minutes, a cron pulls snapshots from 7 cameras between the SR-209 intersection and Alta, counts vehicles on-device (Apple Silicon GPU), and appends a time series — a canyon "busy-ness index" that feeds powder-day traffic logic (anyone who has sat in the red snake on a powder morning knows).

• Cameras: UDOT publishes ~1-minute snapshots at https://www.udottraffic.utah.gov/map/Cctv/<id>. No API key needed; the developer API only adds metadata, not faster frames.
• Model: rfdetr-base (Apache 2.0), run locally — no cloud calls, no per-frame cost. COCO classes filtered to car/truck/bus/motorcycle.
• Dedup: DETR-family models can emit the same physical vehicle under two classes (car 0.40 + truck 0.46). Class-agnostic NMS (detections.with_nms(threshold=0.5, class_agnostic=True)) keeps one box per object — without it, counts run ~20% hot. Found by hand-counting a frame against the model output.
• Guards: near-black frames (night) and dead feeds are logged as such rather than counted as zero traffic.

pip install inference supervision python-dotenv
echo "ROBOFLOW_API_KEY=<your key>" > .env
python canyon_watch.py

Outputs canyon_watch/readings.jsonl (one line per camera per poll, plus a _canyon_total) and annotated snapshots per run.

Cron (every 10 minutes):

*/10 * * * * /path/to/python /path/to/canyon_watch.py >> canyon_watch/cron.log 2>&1

Native (non-Docker) inference on M-series Macs currently crashes with AttributeError: module 'torch.mps' has no attribute 'current_device' — see roboflow/inference#2757. The script includes the one-line workaround.