Traditional city ANPR cameras operate as disconnected islands. This system solves this by operating a 5-stage real-time intelligence pipeline:
1
Edge Detection & Optical Character Recognition (OCR)
YOLOv8-Nano plate detection + CRNN/Transformer OCR running at edge nodes with super-resolution filtering to achieve >90% precision in poor illumination, high angles, and dirt.
2
High-Throughput Ingestion Queue (Kafka / MQTT)
Camera nodes stream structured JSON telemetry payloads: camera ID, GPS, millisecond timestamp, OCR plate string, and feature embedding vector with sub-50ms latency.
3
Plate Normalization & Fuzzy Deduplication
State code parsing (e.g. TN, DL, MH) and Levenshtein distance matching resolve common OCR optical swaps (e.g. '8' vs 'B', '0' vs 'O') across varying lighting.
4
Spatio-Temporal Trajectory Reconstruction Engine
Chaining algorithm links sightings sequentially, filters geographically impossible teleportation speeds, calculates inter-camera segment speeds, and constructs continuous GIS polylines.
5
Macro Urban Analytics & Watchlist Alert Daemon
Aggregates vehicle trajectories into corridor travel times, O-D density matrices, and bottleneck forecasts while continuously matching active police hotlists.