TrafficOP watches junction footage, flags traffic violations with a YOLOv11 vision stack, reads number plates, and produces annotated evidence — with a planned second-stage LLM reviewer that confirms every flag before a challan is issued. Pick a camera below and watch the pipeline run, frame by frame.
This is a demo run through our engine — both clips above were processed by TrafficOP, and the annotated video, detections and evidence shown below are the actual output of that run.
The annotated evidence feed plays here. Each time the vision stack flags a violation, it drops into the incident feed on the right and is written to the database immediately — marked PENDING until Stage 2 reviews it.
This clip is a demo run through our TrafficOP engine.
Vision detectors are fast but produce false positives — shadows, caps, pedestrians behind a bike. Every flagged incident is handed to a vision-language model that looks at the cropped evidence and returns CONFIRMED or FALSE POSITIVE. Only confirmed cases auto-issue a challan; the rest route to a human officer.
Not implemented yet. The Stage-2 reviewer shown here is a demonstration of the intended flow — it runs as a deterministic mock and is not yet wired to a live LLM. Connecting a real Vision-Language Model (and using it to read plates that OCR misses) is the next step — see the roadmap below.
Decode frames from the camera feed.
UVH-26 + helmet + person YOLO models.
No-helmet & triple-riding logic with IoU.
OCR the number plate region.
VLM reviews each crop → verdict.
Confirmed → auto-challan; else officer.
outputs/violations.db
| # | Type | Plate | Det conf | LLM verdict | Reason | Status |
|---|---|---|---|---|---|---|
| — run a camera to populate — | ||||||
Frame sampling, contrast/denoise, low-light & motion-blur handling.
Task 1UVH-26 localises 14 Indian vehicle classes plus riders and pedestrians.
Task 2No-helmet & triple-riding today; extensible to seatbelt, wrong-side, red-light.
Task 3Every flag categorised into a class with a detector confidence score.
Task 4Plate detection + OCR in Indian format (KA 01 AB 1234).
Task 5Annotated evidence video + cropped frames, timestamped per incident.
Task 6Searchable SQLite, per-type stats and the live dashboard above.
Task 7Accuracy, Precision, Recall, F1, mAP — plus throughput in fps.
Task 8Our edge: a VLM will double-check every detection to cut false positives. Demonstrated as a mock today; live model planned.
PlannedThis prototype delivers the full end-to-end flow for two violation types. The remaining items from the brief are planned in both the engine and this console:
Wire the second-stage reviewer to a real Vision-Language Model — today it runs as a demonstration mock.
NextWhen OCR can't read a plate, the LLM will read the registration number from the evidence crop.
NextPlate-OCR accuracy is still limited; improve the preprocessing and recognition model for cleaner reads.
NextCalibrate and refine the per-detection confidence scoring.
NextSeatbelt, wrong-side driving, stop-line, red-light and illegal parking.
NextAccuracy, Precision, Recall, F1 and mAP, plus throughput & scalability.
Next