VEIL-AI

Vehicle Edge Intelligence Logger - advanced AI for vehicular security and monitoring.

Plate detection model

Download a license-plate YOLO checkpoint (for example keremberke/yolov8n-license-plate) and place it at models/yolov8n-license-plate.pt, or set the PLATE_MODEL_PATH environment variable to your custom .pt file before running main.py or main_video.py.

Training with the Indian Kaggle dataset

You can fine-tune a YOLO detector on the Indian license plates with labels dataset directly from this repo:

1. Prepare the splits

   C:/btech/development/VEIL-AI/veilenv/Scripts/python.exe scripts/prepare_indian_lp_dataset.py

   The script downloads the dataset via kagglehub, creates deterministic train/val/test splits under data/indian_lp/, mirrors the original download to data/indian_lp/raw/, and writes both a YOLO data.yaml and a summary.json with useful counts.

2. Train with Ultralytics YOLO

   # Option A: call Ultralytics directly
   yolo detect train model=yolov8n.pt data=data/indian_lp/data.yaml epochs=80 imgsz=640
   
   # Option B: use the helper wrapper (adds AdamW + cosine LR + patience)
   C:/btech/development/VEIL-AI/veilenv/Scripts/python.exe scripts/train_indian_lp.py \
   	 --model yolov8n.pt \
   	 --epochs 80 \
   	 --imgsz 640

   Replace yolov8n.pt with any starting checkpoint. Final weights will be emitted under runs/detect/.../weights/best.pt.

3. Validate + wire into VEIL

   # Quick validation report on the held-out test split
   yolo detect val model=runs/detect/<run-name>/weights/best.pt data=data/indian_lp/data.yaml split=test
   # or use the helper if PowerShell quoting is painful
   C:/btech/development/VEIL-AI/veilenv/Scripts/python.exe scripts/val_indian_lp.py \
   	--model runs/detect/<run-name>/weights/best.pt \
   	--split test --imgsz 640 --batch 16
   
   # Update VEIL to use the refined detector
   copy runs/detect/<run-name>/weights/best.pt models/yolov8n-license-plate.pt
   # or set PLATE_MODEL_PATH directly in config.py / env vars

   The latest YOLOv8n fine-tune (test split with 202 images) yields precision 0.995, recall 0.990, mAP50 0.995, and mAP50-95 0.865, confirming the new checkpoint surpasses the base detector.

   Afterwards rerun python main_video.py (or the live camera pipeline) and compare plate detection accuracy + speed.

Batch-testing detection + OCR on a labeled dataset

To measure end-to-end accuracy on any folder of plate images:

python scripts/eval_plate_dataset.py \
	--images path/to/plate/images \
	--labels path/to/labels.csv \
	--output reports/plate_eval.csv

The label file can be a CSV (image,plate columns) or JSON with the same keys. Add --fallback-stem if filenames already encode the ground truth text. The script reports detection hit rate, OCR exact-match rate, average similarity, and optionally writes a per-image CSV so you can inspect failures quickly.

OCR engine configuration

• Default (Windows accuracy): EasyOCR (PyTorch-based) remains the most reliable reader out of the box. Keep OCR_ENGINE=easyocr (default) for best results while you iterate on tuning.
• Optional lightweight mode: set OCR_ENGINE=tesseract if you install pytesseract plus the Tesseract binary (winget install tesseract-ocr.tesseract on Windows or sudo apt install tesseract-ocr on Debian/Raspbian). Useful when you need a Pi-friendly footprint.
• Additional knobs for Tesseract: TESSERACT_CMD=C:/Program Files/Tesseract-OCR/tesseract.exe, TESSERACT_LANG, TESSERACT_PSM, TESSERACT_OEM.
• Regardless of the engine, regex-enforced post-processing only accepts valid Indian plate formats, so uncertain reads are dropped rather than logged as hallucinations.

Cloud sync configuration

By default the pipeline attempts to sync completed entries to Firebase Cloud Firestore. Provide credentials in one of two ways:

1. File path – download your Firebase service-account JSON and set FIREBASE_CREDENTIALS to its absolute path (defaults to serviceAccount.json in the project root).
2. Inline JSON – set FIREBASE_CREDENTIALS_JSON to the raw JSON string (useful for CI or secrets managers). The app writes it to .cache/firebase_credentials.json automatically.

If you prefer a REST endpoint instead of Firebase, set CLOUD_PROVIDER=rest, point CLOUD_ENDPOINT to your API, and set CLOUD_API_KEY to the corresponding bearer token. Use --no-cloud on main.py / main_video.py for fully offline runs.