Automating License Plate Recognition for a Traffic Enforcement System

Industry: Public sector, road safety

Project type: Computer vision, image classification, legacy system integration

The Problem

An automated speed enforcement system ran on classical computer vision: pixel-based edge detection for license plates, Haar cascade–style algorithms — no neural networks. The logic was linear: a camera clocks a speeding vehicle → captures a photo → the plate is read automatically → the citation moves into processing.

The pipeline broke down in three predictable ways. Snow, dirt, and glare degraded image quality at the camera level. A share of plates were deliberately obscured — taped or painted over — specifically to defeat the reader. And critically: any plate that wasn't read with full confidence couldn't move forward in the pipeline at all. Every partial read fell into a manual review queue.

That queue was staffed by roughly 300 people working in shifts. This wasn't a one-time bottleneck — it was a standing, growing line item. Every new camera added to the network meant more volume into the same queue.

The Solution

The insight: an unreadable plate still leaves the system with useful signal — the photo of the vehicle itself. Make and model are visible in the frame even when the plate is partially obscured.

Make/model classification. A convolutional neural network was trained to identify vehicle make and model from the camera frame — roughly 100 of the most common brands in the country, plus a long tail of less common ones.

Cross-referencing against the vehicle registry. Combining the predicted make/model with whatever plate digits were legible, the system queried the vehicle registration database and visually matched candidates against the captured frame.

Training data without manual labeling from zero. Rather than hand-labeling a dataset from scratch, the team reused data the system already had: any photo where the plate had been read successfully by the camera was automatically treated as labeled — a correctly read plate meant make and model were already known from the registry. Manual labeling was reserved for edge cases and rejected reads only, which cut both the volume and the cost of labeling dramatically.

Route reconstruction. A separate module: given proper legal authorization, the system could reconstruct a specific vehicle's movement across the city over a defined window — stitching together its appearances across the camera network into a timestamped sequence with stop points.

Rollout and Organizational Resistance

The first model iteration was tuned conservatively — a high confidence threshold before the system would act automatically. Once it became clear that even at that cautious threshold the system was absorbing meaningful volume, the managers overseeing the manual review team grew concerned about their team's future. The threshold was deliberately lowered so automation would take on more of the queue and make the case for where manual review was no longer needed.

This is the standard dynamic in any deployment that displaces manual labor: the resistance rarely comes from the technology. It comes from the people whose jobs are changing.

Results

• Automated processing share increased by roughly 40% in the first iteration.
• Confidence threshold was deliberately lowered from ~95% to ~80% to shift more volume out of manual review.
• A significant portion of the 300-person review staff became unnecessary for this class of case — an exact reduction figure wasn't tracked.
• Delivery timeline: ~6 months, including integration and documentation.

Team

15 people: 2 backend engineers, 2 full-stack engineers, 2–3 ML engineers on model training, 2–3 analysts, 1 QA, 1 interface designer, 1 product manager.

Economics: Delivery Cost vs. Client Price

For a new B2G engagement — a city of roughly 15 million people, ~1,500 cameras, with an existing labeled dataset — pricing lands at approximately €500,000.

What Would Change If Built Today

Rebuilding this with current AI tooling changes the economics in two places.

Backend and frontend development compress substantially — most of this class of engineering problem is already solved in open-source repositories, and AI coding assistants scaffold a working system in the time it used to take to write it by hand from a blank file.

Training the classification model — previously ~2 months of work for an experienced specialist — realistically compresses to 2–3 weeks, driven by pretrained foundation models and AI-assisted tooling for writing labeling and data-extraction scripts.

Net estimate: timeline roughly halves — 6 months down to 3 — and budget drops 20–30%.