Automatic Number Plate Recognition (ANPR) – Real-Time AI Video Analytics

Project Summary

This project implements a production-grade Automatic Number Plate Recognition (ANPR) system for real-time vehicle monitoring at entry and exit points.
The system is designed for 24×7 continuous operation using GPU-accelerated video analytics and is suitable for industrial plants, parking facilities, and gated premises.

The pipeline processes live RTSP camera streams, performs vehicle and license plate detection using NVIDIA DeepStream, and applies OCR-based plate recognition with PaddleOCR, achieving high accuracy even under challenging real-world conditions.

---

Key Metrics & Performance

• Inference Speed: 25–30 FPS per camera stream
• OCR Accuracy: Improved from ~72% → 92–95%
• False Positives Reduction: ~65–75%
• End-to-End Latency: ~150–200 ms per vehicle event
• Vehicle Speed Handling: Reliable up to ~30–40 km/h
• Deployment Mode: Real-time, multi-stream, GPU-accelerated

---

Project Overview

The system ingests live RTSP streams and performs:

1. Vehicle & License Plate Detection using NVIDIA DeepStream
2. Dynamic cropping of detected license plates
3. OCR text extraction using PaddleOCR
4. Event logging with metadata:
   • Recognized plate number
   • Timestamp
   • Camera location
   • OCR confidence score
   • Snapshot image

All vehicle events are stored in a structured, audit-ready format, enabling:

• Security monitoring
• Automated access control
• Compliance and forensic auditing

---

Challenge: PaddleOCR in Real-World Conditions

While PaddleOCR performed well on clean images, significant challenges were observed in live streaming environments:

• Night-time glare & uneven illumination
• Rain reflections and low-contrast plates
• Motion blur from fast-moving vehicles
• Angled camera views & non-standard fonts

Initial Issues Observed

• Character-level accuracy dropped to ~70–75%
• Duplicate OCR results across consecutive frames
• Partial reads and low-confidence outputs cluttered logs

---

Solution Approach & Technical Optimizations

To stabilise OCR performance, a multi-stage preprocessing and intelligence layer was implemented.

🔹 Pre-processing Enhancements

• CLAHE-based contrast enhancement
• Noise reduction & glare suppression
• Dynamic plate cropping from detection outputs
• Perspective correction for skewed or angled plates

🔹 Post-processing & Intelligence Layer

• Confidence thresholding to discard unreliable OCR outputs
• Regex-based validation and character cleanup
• Similarity matching to suppress near-duplicate reads
• Temporal filtering to trigger OCR only on best-quality frames

---

Measured Improvements (Post-Optimization)

Metric	Before	After
OCR Accuracy	~72%	92–95%
False Positives	High	↓ 65–75%
OCR Confidence Stability	Low	↑ ~40%
FPS per Stream	25–30	Maintained
Latency	~200 ms	150–200 ms

---

Final Solution Delivered

The final system is a robust, scalable, and production-ready ANPR pipeline capable of operating reliably under:

• Low-light and night-time conditions
• Rain and reflective environments
• Motion blur and angled camera placements

It produces clean, duplicate-free vehicle logs with full traceability:

• Timestamped events
• Camera identifiers
• OCR confidence scores
• Image evidence

This makes the solution suitable for security auditing, automated access control, and compliance tracking.

I have passed parameters to paddleocer to manage and text angle.

🔄 End-to-End Pipeline (High-Level Flow)

RTSP Camera Stream
        ↓
DeepStream Vehicle & Plate Detection (25–30 FPS)
        ↓
Dynamic Plate Cropping
        ↓
Image Preprocessing
(CLAHE, Denoising, Glare Removal, Angle Correction)
        ↓
PaddleOCR Text Recognition (92–95% Accuracy)
        ↓
Post-processing
(Confidence Filtering, Regex Cleanup, Duplicate Suppression)
        ↓
Structured Logging + Snapshot Storage