Vehicle & Plate Detection from the Cameras You Already Have

One platform for traffic detection and license plate reading. Count vehicles, catch wrong-way drivers, flag emergency vehicles, spot parked vehicles, and read plates, all using your existing surveillance cameras.

Works with WINK Media Router to add detection and plate reading to cameras you already have.

Request a Demo See How It Works

California license plate reading 8WINK22

What You Get

Vehicle Counting & Classification
Wrong-Way Driver Detection
Emergency Vehicle Alerts
Parked & Stopped Vehicle Detection
License Plate Recognition
Multi-Camera Vehicle Tracking
Hot List & Watch List Alerts
Journey Time & Route Analysis

Overview

What WINK Traffic and LPR Is

Vehicle detection and plate reading that runs on the cameras an agency already owns, rather than on a separate network of dedicated ALPR hardware.

The conventional approach to plate reading starts with procurement. Dedicated ALPR cameras go in at chosen collection points, aimed tightly, lit deliberately, and priced per location. The result works well and covers the handful of sites the budget stretched to, which means the coverage map is decided by cost rather than by where vehicles actually go.

This product starts from the opposite end. A city or a DOT typically operates hundreds of cameras already, mounted on poles, pointed at intersections and corridors, with lighting nobody controls. Reading plates from that material is harder, and we accept the loss on individual frames in exchange for something the dedicated approach cannot offer: every camera you own becomes a collection point.

It is an add-on to WINK Media Router, so the analytics run against streams already flowing through the platform. A camera that is in the system can be turned on for detection without new cabling, a new network, or a second video path.

At a Glance

Runs on existing cameras, not dedicated ALPR hardware
Vehicle counting and classification by camera profile
Wrong way, stopped, and emergency vehicle detection
Plate reading with regional model libraries
Multi-camera association for routes and journey times
Edge or cloud processing, mixed in one deployment
Full audit trail with source frames retained per event

Many Cameras, One System

A vehicle that passes one camera and then another tells you a lot more than either camera tells you alone. We tie the observations together.

Connect dozens or hundreds of cameras, whether intersections, highway corridors, or toll plazas, into one place. Each camera can run a different detection profile: highway cameras watch speed and classification, intersection cameras watch pedestrians, toll cameras read plates.

When a vehicle leaves one camera's view and shows up on the next, the system stitches the two observations together. No special hardware, no manual calibration.

Run the heavy lifting at the edge for cameras that need instant alerts. Run it in the cloud for everything else. Mix both in the same deployment. You don't have to commit to one model.

Coverage compounds. The more cameras you point at the problem, the more useful the picture gets.

Where Did That Car Come From, and Where Did It Go?

Once we've read a plate at one camera and read it again at another, we know the route, the time it took, and the volume on that path. Multiply that across thousands of vehicles and you can answer real questions: which corridors are slowing down, where demand is shifting, which routes drivers actually pick. No GPS feeds, no probe data, no surveys, just the cameras you already operate.

The corridor travel times that come out of this can feed straight into 511 systems and dynamic message signs.

The Plate Is the Signature

No magic. We read the plate, attach an identifier to the vehicle, and track that identifier across every camera that sees it. Routes, dwell times, repeat visits, hot-list matches. They all come from that one observation, done well, at scale.

Built for Real-World Cameras, Not Lab Conditions

Most LPR products only work on dedicated, perfectly-mounted ALPR cameras with a tight angle, controlled lighting, and a narrow field of view. Ours doesn't make that demand. Plug in the city cameras you already operate, pole-mounted, off-axis, glare-prone, IR at night, wide-area pan-and-tilt, and we'll catch what we can.

A Model Tuned to Your Cameras

This is why we get usable LPR out of cameras other vendors give up on: we train a model specifically for your camera makes and models, your optics, your lighting conditions, your mounting angles. We learn the way your install distorts plates and account for it. You'll have higher success on imperfect cameras than off-the-shelf ALPR products can deliver, because no one else is tuning to your exact equipment.

Turn Your Existing Network Into an LPR Mesh

That's the part that didn't used to be possible. Until now, building a region-wide LPR mesh meant installing dedicated ALPR cameras at every collection point, an expense most agencies could never justify. By making LPR work on the cameras you already have, every existing camera becomes a node in the mesh. Suddenly your whole city is the collection grid.

A Library of Pretrained Plate Models

Different regions have different plate formats. We maintain a large library of pretrained models covering plate formats from across the Americas, Europe, and Asia. When your region or use case needs something different, we can quickly train a model with you using your own footage.

It Gets Better Over Time

This isn't a one-and-done install. We retrain the models against anonymized data from your deployment, so recognition on your specific cameras gets sharper as we go. You're not locked into the accuracy you bought on day one.

Hot Lists and Search

Load watch lists for stolen vehicles, AMBER alerts, or agency BOLOs and you get an alert the moment that plate shows up anywhere on the network. Search any plate at any time across the full history.

Plays Nicely With What You Already Use

Detections show up in your ATMS, CAD, video wall, or message signs, not in a separate window your operators have to remember to check.

Push events through webhooks, REST API, email, or SMS. When a wrong-way driver triggers an alert, the right person gets paged on the right channel within seconds. Set up escalation chains so an unacknowledged alert moves up the ladder automatically.

If you run WINK Media Router, every camera already in the system can be turned on for analytics with one toggle. WINK Wall automatically pops up the relevant camera with the detection overlay so the operator doesn't have to hunt for it.

What It Watches For

A short list of the things people most often ask about. There's more, so ask us.

Vehicle Counting & Classification

Count and classify all 13 FHWA vehicle types per lane and direction. Hourly, daily, and weekly volume reports for traffic studies and capacity planning.

Wrong-Way Driver Detection

Instant alerts when a vehicle is detected traveling against traffic. Critical for highway entrance ramps, divided arterials, and one-way street networks.

Emergency Vehicle Detection

Automatic identification of fire, EMS, and police vehicles. Route preemption signals to traffic management systems and clear the path automatically.

Parked & Stopped Vehicle Detection

Identify vehicles parked illegally, stopped on shoulders, or abandoned in travel lanes. Configurable dwell-time thresholds reduce false positives.

License Plate Recognition

Real-time plate capture from existing cameras with hot list alerts, plate search, and full chain-of-custody logging for evidentiary use.

Speed & Flow Analysis

Average speed estimation, queue length measurement, and traffic density tracking. Detect congestion as it forms, not after it's reported.

Pedestrian & Cyclist Detection

Identify vulnerable road users at intersections and on highways. Trigger warnings, signal extensions, or alerts when pedestrians enter unsafe zones.

Weather & Visibility

Detect rain, snow, fog, and reduced visibility conditions. Integrate with dynamic message signs and variable speed limit systems.

Multi-Camera Tracking

Follow vehicles across your entire camera network. Reconstruct travel paths, measure corridor times, and answer movement questions.

Who's Using This

State and Municipal DOTs

Centralized traffic intelligence across thousands of cameras. Real-time corridor monitoring, automated incident detection, and travel time data feeding 511 systems and dynamic message signs.

Law Enforcement

Plate recognition with hot list alerts for stolen vehicles, AMBER alerts, and BOLOs. Vehicle search across the camera network with full audit trails for investigations.

Toll Authorities

Plate-based tolling support, queue management, and vehicle classification for dynamic pricing. Reduce wait times and increase throughput at toll plazas.

Emergency Management

Real-time situational awareness during incidents. Automatic incident detection, evacuation route monitoring, and multi-agency coordination through video wall integration.

The Software

What Operators Actually See

Screens from a running deployment rather than mockups.

WINK LPR dashboard showing per camera detection counts, average confidence, an hourly traffic heatmap, and a searchable detections table — **Live dashboard**Per camera detection and unique vehicle counts with average and high confidence rates, an hourly heatmap against the fourteen day average, cross camera pairing with timing delta, and a searchable detection log.

Detection settings screen showing the pipeline from camera frame through region of interest crop and YOLO plate detection to OCR voting — **Detection pipeline**The path a frame takes, drawn with the numbers that matter: region of interest crop, model input size against plate pixel width, inference time per model, and the OCR voting stage at the end.

Capabilities screen listing real time ALPR, NPU accelerated inference, custom OCR engine, vehicle classification, direction intelligence, registry integration, training pipeline, multi camera fusion, and platform integration — **Capabilities and architecture**What the deployment is running, from YOLO v11 detection and the CTC-CRNN OCR engine to NPU offload, multi camera fusion, and the continuous training pipeline, with the component stack underneath.

System monitoring screen showing host status, temperature, CPU and memory graphs, and Hailo NPU inference latency per port — **System monitoring**Host and NPU state side by side: service status and uptime, temperature, CPU and memory, and per port inference latency and tracker FPS from the Hailo accelerator, so the people running it can see what it is doing without opening a shell.

Technical Brief

Why We Vote Across Many Crops Instead of Enhancing One

WINK Streaming · Computer vision in the field

Conventional plate recognition needs a dedicated camera: low, close to the lane, aimed along the direction of travel, infrared lit, short exposure. Do all that and the reader is handed a near ideal image. The weakness is not technical, it is that this costs money per location, so agencies get six collection points in a city while hundreds of general purpose cameras watch vehicles all day and contribute nothing.

Super resolution is the obvious fix, and it does not work

We tested it against production data. Three pipelines, including a 1.21M parameter Real-ESRGAN model trained by Tencent ARC on millions of images, all returned 0.0% exact match and 0.4% character accuracy, at 126ms per crop against 8.9ms for the reader itself. Below roughly 100 pixels of plate width the information is not in the input, and a model that produces a plausible character is producing a hallucination, which is worse than a blank result for anything that may end up in evidence.

Looking more than once does work

A vehicle passing a camera generates 15 to 20 crops that fail in different ways. Reading all of them and voting across the results returned 92.0% exact match over a thousand vehicles, against 79.5% for the best single crop and 77.4% for fusing crops into one composite. Where voting and fusion disagreed, voting was right 156 times and fusion 10. Good crops carry the vote and bad crops are noise, upscaled or not.

In practice: corridor journey times without probe data

A plate read at one camera and again at another gives a route and an elapsed time. Repeated across thousands of vehicles a day on cameras the agency already owns, that produces corridor performance data with no GPS feeds, no probe data purchase, and no new roadside hardware. Individual read rates on pole mounted cameras are lower than a dedicated ALPR lane; the network level answer is better, because there are far more collection points.

The full experiment, including the crop size distribution across 314,979 production crops over three months, is in Neural Super Resolution as an LPR Pre-Filter. Detection runs as an add-on to WINK Media Router, with WINK Crossroad for monitoring.