TTrafficOP
PIPELINE claude-opus-4-8
STANDBY
Gridlock Hackathon 2.0 · Computer Vision + LLM Enforcement

One camera feed.
Every violation, caught and confirmed.

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.

STAGE 1 · UVH-26 YOLOv11-S vehicle + helmet + person detection STAGE 2 · LLM reviewer confirms / rejects each flag (planned — mock today) STORE · SQLite, written live

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.

01

Live detection console

select a camera to begin

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.

▷ No feed loaded.
Choose a camera above and press RUN PIPELINE.
SCANNING — STAGE 1 VISION
FRAME 0 · T 0.0s
idle 0 incidents flagged

This clip is a demo run through our TrafficOP engine.

⚠ Incident feed 0
Incidents appear here as the
vision stack flags them.
02

Stage 2 — LLM confirmation

runs after the clip ends

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.

00

Ingest

Decode frames from the camera feed.

01

Detect

UVH-26 + helmet + person YOLO models.

02

Flag

No-helmet & triple-riding logic with IoU.

03

Read plate

OCR the number plate region.

04
★ LLM

Confirm

VLM reviews each crop → verdict.

05

Challan

Confirmed → auto-challan; else officer.

03

Live dashboard

read straight from outputs/violations.db
Violations detected
0
LLM-confirmed
0
Flagged / pending
0
Camera runs
0

Violations by type

no data yet

Stage-2 verdict split

0%confirmed
Confirmed 0
False positive 0
Pending 0

Violation records

#TypePlateDet conf LLM verdictReasonStatus
— run a camera to populate —
04

Mapped to the brief

done today · more in progress
🌦️

Image preprocessing

Frame sampling, contrast/denoise, low-light & motion-blur handling.

Task 1
🚗

Vehicle & road-user detection

UVH-26 localises 14 Indian vehicle classes plus riders and pedestrians.

Task 2
🚨

Violation detection

No-helmet & triple-riding today; extensible to seatbelt, wrong-side, red-light.

Task 3
🏷️

Classification + confidence

Every flag categorised into a class with a detector confidence score.

Task 4
🔢

License-plate recognition

Plate detection + OCR in Indian format (KA 01 AB 1234).

Task 5
🖼️

Evidence generation

Annotated evidence video + cropped frames, timestamped per incident.

Task 6
📊

Analytics & reporting

Searchable SQLite, per-type stats and the live dashboard above.

Task 7
📈

Performance evaluation

Accuracy, Precision, Recall, F1, mAP — plus throughput in fps.

Task 8
🧠

2nd-stage LLM reviewer

Our edge: a VLM will double-check every detection to cut false positives. Demonstrated as a mock today; live model planned.

Planned
05

Future implementation

what we're building next

This 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:

Live Stage-2 LLM

Wire the second-stage reviewer to a real Vision-Language Model — today it runs as a demonstration mock.

Next

LLM plate fallback

When OCR can't read a plate, the LLM will read the registration number from the evidence crop.

Next

OCR refinement

Plate-OCR accuracy is still limited; improve the preprocessing and recognition model for cleaner reads.

Next

Confidence refinement

Calibrate and refine the per-detection confidence scoring.

Next

More violation types

Seatbelt, wrong-side driving, stop-line, red-light and illegal parking.

Next

Formal metrics

Accuracy, Precision, Recall, F1 and mAP, plus throughput & scalability.

Next