License Plate Recognition using Image Processing

Project at a Glance

• Client/Team: Thesis Nest
• Domain: Intelligent Transportation, Smart Parking, Security Automation
• What we built: A fast, privacy-aware License Plate Recognition (LPR/ANPR) system that ingests live RTSP/ONVIF camera streams, detects vehicles, localizes plates, and extracts plate text in real time—complete with whitelists/blacklists, alerts, and searchable logs.
• Runs on: Edge (Jetson/IPC) or Cloud GPU; works with existing CCTV/NVR infrastructure.
• Highlights: High accuracy across day/night and weather changes, supports multiple countries’ plate layouts, and offers APIs for parking, access control, and tolling systems.

The Challenge

Parking gates, campuses, industrial sites, and city checkpoints need reliable vehicle identification without manual checks. Common hurdles:

• Low-quality feeds: motion blur, glare, compression artifacts
• Varied plates: different fonts, sizes, multi-line formats
• Operational latency: gates shouldn’t keep drivers waiting
• Privacy & compliance: handling plate data responsibly

Our Approach

A three-stage pipeline optimized for speed and accuracy:

1. Vehicle & Plate Detection
   Lightweight object detectors find vehicles first, then focus on plate regions. This two-pass strategy boosts recall in cluttered scenes.
2. Plate Enhancement & Rectification
   Perspective correction, deblurring hints, denoising, and contrast stretching improve OCR robustness. Adaptive exposure handling manages day/night transitions.
3. OCR & Post-Processing
   OCR reads alphanumerics with country-specific character sets. Lexicon rules, checksum patterns, and confidence heuristics clean up results. Final outputs include plate text, confidence, and plate ROI for auditing.

Edge Intelligence: Optional on-camera/edge inference keeps video on-prem; only metadata leaves the site if desired.

Key Features

• Real-Time Recognition: Sub-second inference per frame on typical deployments
• Country Profiles: Layout rules for single-line, double-line, stacked region codes
• Watchlists: Whitelist/blacklist with instant alerts via SMS/Email/Webhooks
• Event Log & Search: Filter by time, camera, partial plate, confidence, action taken
• Gate Control: REST/WebSocket integrations to trigger barriers or alarms
• Multi-Camera Tracking: Merge sightings to reduce duplicates across lanes
• Operator Console: Live video overlay with plate boxes and confidence badges
• Privacy Controls: On-prem processing, auto-redaction in stored images, retention policies

Results & Impact

• Faster throughput at gates → reduced congestion and better user experience
• Improved security → instant notifications for flagged vehicles
• Operational visibility → reports on peak hours, frequent visitors, dwell times
• Lower TCO → works with existing cameras; edge deployment minimizes bandwidth

(Actual accuracy and latency depend on camera placement, resolution, plate style, and lighting; we run a site survey and pilot to tune parameters.)

What Makes It Different

• Accuracy under real conditions: Robust to motion blur, skew, and headlight glare
• Country-aware OCR: Profiles reduce false positives across diverse plate formats
• Edge-first design: Keep video on-prem; send only the metadata you need
• Developer-friendly: Clean REST/WebSocket APIs + SDK snippets for rapid integration

Deliverables

• Deployed LPR service (edge or cloud)
• Operator dashboard with live overlay, logs, and reports
• API/SDK + integration guide (parking, access control, ERP)
• Site survey report & camera placement recommendations
• Admin/DevOps playbook and MLOps handover
• Optional fine-tuning on your cameras and plate types