The Mistakes Made by License Plate Reader Cameras

Automated License Plate Reader (ALPR) cameras make mistakes. This fact has been grabbed on by and exploited by critics of the technology.

The vast majority of these critics have next to no idea what they're talking about. They read about "mis-scans" or "misreads" or some other bad "hit" or alert. They attach this to some sympathetic news story of a bad traffic stop. Then they use this as a justification to abandon ineffective technology. 

When I ask these critics about the effectiveness measures, or statistical methodology, they take one (1) of two (2) responses:

1. ignore me;
2. attack me, but still ignore the questions.

As a long-time user of ALPR tech for almost two (2) decades, I'm fairly aware of the mistakes that the system makes. As am I aware of the mistakes humans make when engaging with ALPR data.

Allow me articulate exactly what mistakes, errors, or inaccuracies occur with this tech:

• Character Misidentification. This is when the system reads a Q instead of the D.
• State Misidentification. This is when an Illinois plate is categorized as Connecticut.
• Type Misidentification. This is when a motorcycle plate is categorized as a Passenger plate.
• Year Misidentification. Inability to decipher 2025 from 2026 issuance; especially for fleet or dealer plates. 

These are completely expected. Why?

• Cars move fast.
• Lighting is varied.
• Weather - rain, snow.
• Obstructions - bike racks, objects sticking out of trunk, trailer hitches.
• Dirty plates.
• Worn, faded, pealing plates.
• Lots of states.
• Lots of plate types within each state (trucks, rental, handicapped, trailers, cars, specialty, etc).
• Variety of font, styles of letters/numbers, stacked digits, etc.

Humans need to verify these issues. For each and every hit analyzed, And for each and every historical read analyzed. 

But there are other issues with hotlists too.

• Police agencies don't properly or immediately clear Stolen or wanted plates from state/national databases, after they've been recovered.
• Most ALPR systems only "clear" the automated hotlists once per twelve (12) hours, or on some other cycle.
• Mistakes might be made by those entering plates into state/national databases, in terms of state, year, types, or digits. Or model or color of car.
• Victims of auto theft often self-recover their cars by themselves, thereby skipping the database removal step.
• Reported stolen plates might be recycled by the original owner on a new car, or otherwise reissued by state organization.
• Cops not removing/purging plates from manual hotlists when the cars are no longer wanted; or cops not putting in clear, precise narratives in the entry's comments/notes fields.

These above issues with the state/national databases have existed since these databases have been used. They are mistakes made by humans, or reflecting back on delays or slow workflows. Cops have been used to these issues for 50 years. Some of the manual entry lists are newer; but cops have all been trained on these issues.

So when a critic cites some sort of error rate or statistic on bad traffic stops due to this technology, I tend to ask what occurred and which of the above bullet points was at play. Or combination of the bullet points.

Because some of the errors call into question the reliability, consistency, or accuracy of the technology. Other errors highlight human mistakes in data entry. Other errors call out poor judgement on the part of the street cops. 

For me, to expect near-perfect reliability, consistency, or accuracy of a camera and its software to be able to recognize characters, states, years, and types under the real-life conditions -- that's pretty unrealistic at this point. And I think we all know that this tech is not much more than a wink-and-nod in the direction of a possible match. It's not perfect, and none of us expect it to be.

We also expect there to be mistakes in the data entry side. There will always be mistakes. 

It's up to humans-in-the-loop to verify, check, double-check, confirm each and every scan and hit by an ALPR system. It's not rocket science. There are established procedures to do just that. 

But they don't always catch every possible vulnerability in the chain-of-events.

Owners driving their recently recovered stolen auto will be stopped.

Owners who had a license plate stolen will be stopped.

Drivers borrowing cars will be stopped for some combination of mistakes.

While we try to minimize these mistakes, they're going to happen.

And when drivers or occupants want to resist, argue, flee, debate, or complicate things for the cops during an attempted "bad" traffic stop .... even worse things will happen. FAFO. 

In summary, the ALPR technology makes mistakes. It's still extremely valuable at giving cops a direction in which to look and further investigate. Just because the systems aren't perfect doesn't make them ineffective. 

Also, cops make mistakes. That's on us to fix these behaviors, workflows, processes, and procedures.

If we look at ALPR as a helper, and not the sole authority, it's extremely effective at what it's designed to do. 

Even when the camera makes a mistake. 

***

The purpose of this article is for technological and workflow mistakes that give ALPR a bad reputation. This has nothing to do with privacy, Big Brother, surveillance state, or any other arguments on the tech. Please keep comments limited to the topic at hand. Thanks. 

***

Lou Hayes, Jr. is a detective supervisor in a suburban Chicago police department, collaterally detailed to a regional major crimes (homicide) task force.  He has a passion for multi-jurisdictional crime patterns, criminal networks, & regional intelligence. With a background in training, he studies human performance, decision-making, creativity, emotional intelligence, & adaptability. 

Follow Lou on LinkedIn, & also the LinkedIn page for The Illinois Model. ***