BesirVelioglu/id-card-detection-rd-report

ID Card Detection, Tracking & Re-Identification — R&D Report & Prototype

Overview

This repository contains a comprehensive R&D investigation into the best modern approaches for real-time ID card detection, tracking, re-identification, and reference matching from mobile/video input, plus a working prototype implementation.

Contents

• RD_REPORT.md — Full R&D report (50+ pages) covering:

  • Current system critique and limitations
  • Latest research findings (30+ papers, 2022-2026)
  • Dataset evaluation table
  • 3 architecture options (Edge / Production / Research-grade)
  • Final architecture recommendation
  • Experiment design and metrics
  • Synthetic data generation pipeline
  • Training and fine-tuning recipes
  • Implementation roadmap

• enhanced/ — Prototype implementation:

  • config.py — Full configuration system with Edge/Production/Research presets
  • detector.py — YOLOv11 / D-FINE / RT-DETR detector with mock fallback
  • tracker.py — OC-SORT multi-object tracker (replaces ByteTrack)
  • feature_extractor.py — DINOv2 embedding extractor
  • reid_manager.py — EMA gallery for card re-entry detection
  • reference_matcher.py — SuperPoint + LightGlue reference matching
  • quality_scorer.py — Card quality assessment
  • pipeline.py — End-to-end pipeline orchestration

• scripts/ — Evaluation and data generation:

  • evaluate_detection.py — mAP, precision/recall evaluation
  • evaluate_tracking.py — MOTA, IDF1, ID-switch metrics
  • evaluate_reid.py — ROC/AUC, CMC, EER evaluation
  • generate_synthetic_cards.py — Privacy-safe synthetic ID card generator

• tests/ — Integration tests:

  • test_system.py — 68 tests covering all components

Recommended Architecture (Option B: Balanced Production)

Component	Model	HF Model	Latency
Detector	D-FINE-S	Fine-tune from D-FINE GitHub	3.5ms (T4)
Tracker	BoT-SORT-ReID	Custom + DINOv2	~5ms
Re-ID	DINOv2-base	facebook/dinov2-base	~15ms
Reference Match	SuperPoint + LightGlue	ETH-CVG/lightglue_superpoint	sub-10ms
Total			~25ms

Quick Start

from enhanced.config import PipelineConfig
from enhanced.pipeline import IDCardPipeline

config = PipelineConfig.production_config()
pipeline = IDCardPipeline(config)

# Process a frame
result = pipeline.process_frame(frame)
for card in result.cards:
    print(f"Card {card.track_id}: {card.quality_label} (conf={card.confidence:.2f})")

Tests

pip install opencv-python-headless numpy faker
python tests/test_system.py
# Expected: 68/68 passed

Key Research References

1. D-FINE (arxiv:2410.13842) — Detection
2. OC-SORT (arxiv:2203.14360) — Tracking
3. BoT-SORT (arxiv:2206.14651) — ReID Tracking
4. DINOv2 (arxiv:2304.07193) — Instance Embeddings
5. LightGlue (arxiv:2306.13643) — Feature Matching
6. SAM2MOT (arxiv:2504.04519) — Research-grade Tracking
7. DocXPand (arxiv:2407.20662) — Synthetic Data

License

Research prototype — see individual model licenses.