License Plate Recognition & Vehicle Intelligence

Fixed LPR installations at facility entrances, parking structures, and toll plazas must achieve read rates above 95% despite wide variation in plate conditions: faded lettering, non-standard frames, tinted plate covers, temporary tags, and state-to-state format differences across 50+ US plate designs. Mobile LPR deployments on patrol vehicles add motion blur, variable lighting, and off-axis capture angles that reduce accuracy further. Many legacy LPR systems deliver 70-80% read rates in real-world conditions—far below the threshold required for enforcement or access control use cases.

Even when plates are read accurately, isolated reads have limited value without contextual enrichment. A raw plate number becomes actionable only when correlated with hotlist databases (NCIC/NLETS for law enforcement, private registries for repossession and collections), vehicle classification data (make, model, color, body type), and historical visit patterns. Most LPR solutions treat the read as the end product; the real value lies in the intelligence layer built on top of structured read data.

Privacy and data retention regulations vary dramatically by jurisdiction. Some states mandate LPR data deletion within 24 hours unless associated with an active investigation, while federal grants require 5-year retention for certain program data. Agencies need configurable retention policies with cryptographic deletion verification and audit trails that satisfy both civil liberties oversight boards and chain-of-custody requirements for prosecution.

Our LPR pipeline uses a two-stage detection architecture. The first stage runs a vehicle detection model (YOLOv8 fine-tuned on traffic datasets) to localize vehicles and classify body type, color, and approximate make/model. The second stage crops the plate region using perspective-corrected ROI extraction and runs an ALPR-specific OCR model trained on a dataset of over 4 million US plates spanning all 50 states, DC, and common Canadian provinces. This two-stage approach achieves 97.2% character-level accuracy on compliant plates and 93.8% on degraded or obstructed plates in field-validated deployments.

Each plate read generates a structured event containing the plate string, confidence score, jurisdiction classification, vehicle metadata (color, make, model, direction of travel), GPS coordinates, camera ID, and a cropped evidence image. Events are published to a Kafka topic and simultaneously written to a PostgreSQL database with PostGIS extensions for spatial querying. A hotlist correlation service subscribes to the event stream and performs sub-100ms lookups against NCIC/NLETS mirrors, state DMV databases, and configurable local watchlists, generating priority alerts for hit matches.

The analytics dashboard presents heat maps of read activity overlaid on GIS maps, enabling patrol route optimization and coverage gap analysis. Historical pattern analysis identifies vehicles that appear at specific locations repeatedly—useful for investigations, parking enforcement, and facility access auditing. All data access is role-gated with full audit logging, and retention policies are enforced at the database level with cryptographic deletion certificates generated when records age out.

Plate detection begins with RTSP stream ingestion from ONVIF Profile S cameras or dedicated LPR sensors (we support Axis P1445-LE, Bosch MIC IP starlight, and Hanwha Wisenet X series natively). Frames are decoded using hardware-accelerated NVDEC and fed into the vehicle detection model at 15 fps. Detected vehicle ROIs are perspective-corrected using a four-point homography transform estimated from lane geometry calibration, normalizing the plate region to a frontal-parallel view regardless of camera mounting angle (up to 35 degrees off-axis).

The OCR stage uses a CRNN (Convolutional Recurrent Neural Network) architecture with a CTC (Connectionist Temporal Classification) output layer, processing the normalized plate image in a single forward pass without requiring character segmentation. The model outputs a character sequence with per-character confidence scores, and a post-processing layer applies jurisdiction-specific format validation (e.g., California plates follow specific alphanumeric patterns) to reject or flag low-confidence reads. For temporary tags and paper plates, a secondary CNN classifier detects the tag format and routes to a specialized OCR path trained on dealer and transit plate layouts.

Hotlist integration uses a local Redis-backed mirror of NCIC/NLETS data synchronized via CJIS-compliant encrypted channels (TLS 1.3 with FIPS 140-2 validated endpoints). The lookup service performs exact matches, partial matches (for degraded reads), and fuzzy matches (Levenshtein distance ≤ 2) against the hotlist, returning match results with the originating case number and alert priority level. All CJIS data handling complies with the FBI CJIS Security Policy v5.9.1, including encryption at rest (AES-256), advanced authentication, and audit logging of every query.