Flipkart GRiD 6.0 — PS-3 · Bengaluru Traffic Police / ASTraM-compatible
DRISHTI turns any camera — pole CCTV, 4G wireless, even a phone — into an AI traffic-enforcement node. The entire pipeline runs locally on a GPU (demoed on a GTX 1650): no cloud, no per-frame cost, no citizen plate data leaving the edge.
| Violation | How |
|---|---|
| Red-Light Violation | vehicle crosses the stop-line through the line while the signal (HSV) is RED |
| Stop-Line Violation | stops past the line during RED (encroachment) — distinct from running it |
| Wrong-Way Driving | sustained motion opposite the dominant traffic flow |
| Illegal Parking | vehicle stationary inside a drawn no-parking polygon |
| Triple Riding | ≥3 riders sharing one motorcycle box |
| No Helmet | finetuned helmet model on rider crops → without_helmet |
Each flagged vehicle gets super-resolved evidence crops + ANPR (plate read) + timestamp + ERS reliability score.
Distant CCTV plates are low-res and unreadable. Our pipeline reconstructs them:
plate.pt (localize)
→ IBP-MFSR (Iterative Back-Projection multi-frame super-resolution — fuses sub-pixel detail across the burst)
→ Real-ESRGAN (GAN upscaling, general-x4v3)
→ PP-OCRv6 (PaddleOCR, reads multi-row Indian plates)
→ plate-format cleanup (strips the IND strip, validates the format)
The custom method recovers plates that are unreadable in any single frame — here a heavily pixelated capture that a normal OCR returns nothing for becomes legible:
More examples:
Left = the original low-res plate the camera captured. Right = DRISHTI's super-resolved reconstruction + the OCR read. The pipeline reads plates a single frame can't.
We didn't bolt together off-the-shelf weights — we finetuned our own detectors on ~27,000 images / 70,000+ annotations of real CCTV traffic and license plates (incl. Indian plates), on Modal L40S GPUs.
| Model | Classes | Trained on | Result |
|---|---|---|---|
| base | YOLO26-nano | COCO | vehicles, riders, pedestrians, traffic lights |
| plate.pt | license plate | ~14.7k images / 17k boxes | 0.98 mAP@50 |
| helmet.pt | helmet / without_helmet | ~12k images / 55k boxes | real-CCTV helmet compliance |
The whole stack is state-of-the-art (2026):
- YOLO26-nano — newest YOLO, NMS-free, real-time, edge-deployable (Jetson / CCTV NPUs)
- PP-OCRv6 (PaddleOCR, ONNX) — 2026 SOTA OCR for the plate read
- Real-ESRGAN + Iterative Back-Projection MFSR — the forensic super-resolution front-end
- BoT-SORT multi-object tracking
Roadmap: with more compute we unify detection + helmet + plate into a single multi-task YOLO26 (one forward pass per frame → even cheaper at the edge), and expand the plate set with more Indian-specific data.
Base YOLO26-nano — vehicles, riders & pedestrians on a live CCTV junction:
Finetuned helmet & plate detectors:
- Command Center — live KPIs + analytics (recharts)
- Upload & Analyze — drop a clip, draw the stop-line + mark lights, watch live GPU detection + challans, with per-model on/off + confidence sliders
- Violations — officer review queue, evidence gallery, approve/reject
- Live Cameras — connect RTSP/HTTP sources, per-camera calibration
- Model Playground — drop any image, test the 3 models live with adjustable thresholds + super-res toggle
- Hotspot Map — Mappls (MapmyIndia) live violation heatmap of Bengaluru junctions, ranked by density — shows enforcement teams where to deploy
- Frontend: Next.js 16 · React 19 · Tailwind v4 · shadcn/ui · Recharts · true-black UI
- Backend (local): FastAPI · YOLO26-nano (Ultralytics) · BoT-SORT tracking · Real-ESRGAN (spandrel) · PP-OCRv6 (RapidOCR) · OpenCV
- Training: Modal (L40S) — finetune yolo26n on Roboflow datasets
# backend (GPU) — the CV engine lives in ./backend
cd backend && pip install -r requirements.txt && python3 server.py # :8000
# frontend (repo root)
npm install && npm run dev # :3000Open Upload & Analyze → drop a traffic clip → draw the stop-line + mark the signal → Run. For live cameras: Live Cameras → Connect source → calibrate once → Start live detection.
See DEPLOY.md for Vercel + cloud-GPU deployment. The backend needs no API keys;
only MAPPLS_KEY (map) is used client-side — .env is gitignored, nothing secret is committed.