Find more answers here!Nevada lawmakers are expected to draft a bill during the 2027 legislative session that would establish statewide guardrails for the use of Flock cameras. Flock cameras and other automated license plate reader (ALPR) technologies have been in communities for years and are often used by law enforcement as a tool in criminal investigations. In Reno, the Flock cameras have been in use since early 2025, according to news reports. While the Reno Police Department has reported successful use of the technology on its , questions remain among some residents over privacy concerns and misuse of the technology.
As the public debate continues, Assistant Professor Ankita Shukla explains how the technology works, what can affect its accuracy and how engineers test these systems for real-world use.
How does a camera turn a picture of a passing car into searchable information?
A simple way to think about it is that the camera captures an image, but computer vision helps the system interpret what is in the image, and an optical character recognition (OCR) system turns the license plate characters into searchable text. First, a computer-vision model detects the license plate within the larger image. An OCR system then reads the letters and numbers and converts them into text. Depending on the system, computer-vision models may also extract visible details such as the vehicle’s color, make or type.
The plate number and vehicle details can be stored with the time and location where the image was captured. Instead of manually reviewing thousands of photographs, an authorized user can search the structured data for a full or partial plate number or certain vehicle characteristics.
One important distinction is that the system is identifying information about a vehicle. It is not looking at the image and automatically determining who was driving.
How accurate are these systems, and what causes them to make mistakes?
Accuracy depends on how well several steps work together. The camera must first capture a clear enough image, the computer-vision model must correctly locate the license plate, and the OCR system must accurately recognize each character. An error at any stage can affect the final result.
Real-world conditions are also much more challenging than the clean images often used in an academic lab. A vehicle may be moving quickly; the image may be captured at night; or glare, shadows, rain or snow may obscure the plate. The plate itself may be dirty, damaged, partially covered or viewed from an angle.
Even a small mistake, such as confusing an “8” with a “B” or a “0” with an “O,” can lead to an incorrect match. This is why it is important to distinguish between what an AI system suggests and what has actually been verified. An automated match can be useful in an investigation, but it should be checked against the original image, the vehicle details and other available information before drawing a conclusion.
How do engineers train and test ALPR systems to ensure they work reliably in real-world conditions?
For any computer-vision system, the data used to train and test it is extremely important. During training, the model should see a wide variety of examples, not only clearly visible license plates on sunny days. That includes different plate designs, camera angles, distances, vehicle speeds, lighting and weather conditions, as well as partially obscured plates. Training is only one part of the process. Engineers also need to test the model on data it has never seen before to determine whether it can work under new conditions, rather than simply perform well on the examples used to develop it.
I would not rely on one overall accuracy number. A system might perform very well during the day but decline significantly at night, and an average can hide that weakness. Engineers therefore need to evaluate performance under different conditions and understand where the system fails.
Testing should also continue after deployment. Environments, cameras and license plate designs change, and new situations may appear that were not represented in the original training data. For AI systems operating in the real world, reliability is not something we test once: it is something we must continue to evaluate.
About the professor
Ankita Shuka is an assistant professor of artificial intelligence in the College of Engineering. Her research interests include deep learning and machine learning techniques for vision and multimodal data, topological data analysis and geometry-driven approaches for learning. From an application perspective, she focuses on the AI for Science and AI for Social Good movements, specifically targeting wildlife conservation and human health.
