Overcoming plate-reading challenges with (and without) AI

You’ve doubtless noticed them, whether or not your business is using them yet: cameras perched on highways, at toll booths, and in parking garages designed to automatically read vehicle license plates. License Plate Recognition — or LPR — seems simple enough on the surface when it works smoothly. But LPR  systems face a surprising number of difficulties that complicate identifying and reading plates. Let’s dive into some of the most common hurdles that these systems encounter and explore ways that AI can help, as well as some more straightforward means to ensure accurate results. 

One of the biggest obstacles for LPR systems is the sheer variety of physical conditions a license plate can be in. 

Camera choice and placement problems

The angle and distance from which a camera views a plate also play a role. Besides being too high or too far, a camera can be too sharply angled, causing what’s known as perspective distortion. The characters may look stretched or compressed, making it difficult for the system to accurately isolate and identify them. Even cameras placed appropriately might not be able to provide accurate reads if they are super wide angle, or cameras that can pan or tilt, and thus be angled in the wrong direction. Choosing a camera that offers manual controls and excellent low-light performance gets you much closer to reliable plate reads.

Lighting and weather conditions

A system’s ability to “see” is only as good as the lighting. Poor lighting conditions — at night, or in heavy shade  — can make characters hard to distinguish from the background. And glare from direct sunlight or badly placed lights can create bright spots that wash out parts of the plate. Weather matters, too: heavy fog, rain, or snow can make an accurate capture nearly impossible.

Typography trouble: plate design and variety

Beyond environmental factors, license plate design is an ongoing challenge. Plates are not standardized globally; even within a single country, there can be a wide array of designs. In the U.S., readers must account for conventional plates from all 50 states, an ever-growing catalog of specialized plates, and government plates issued for state and Federal government agencies. 

Font and character quirks

Some regions use license plate fonts that are thin or stylized, which can be harder for an electronic eye to read than a bold, blocky font. The spacing between characters can also be inconsistent, and some plates feature stacked or multi-line characters, which require sophisticated algorithms to segment and interpret.

International diversity, global challenges

The diversity of license plates across countries is another hurdle. A system designed for one country might fail completely when faced with a plate from another, which could use different character sets, colors, layouts, and permitted vehicle placement. 

AI’s role in overcoming obstacles

The good news is that technology, particularly AI, is providing powerful solutions. 

First things first: AI can be used to detect the presence of a license plate in an image in the first place, even if it’s not perfectly centered. This allows for a more flexible camera setup and reduces the chance of missing a plate entirely.

AI can be used to preprocess images, automatically adjusting for lighting, glare, and perspective distortion before even attempting to read the characters. This kind of initial processing is especially effective for LPR because the cameras in a typical plate-reading system are in fixed, known positions.

But knowing a plate is there isn’t enough: AI can be trained on massive datasets of plate images, including plates that are dirty, damaged, or viewed from different angles. This training allows the AI to recognize patterns and characters even when they are partially obscured or distorted. Instead of simply looking for a certain pixel pattern,” an AI system can intelligently guess what a character likely is based on its shape, context, and the characters around it.

For example, an AI model can be trained to recognize a “C” even if the bottom half is covered in mud, as long as it has seen countless examples of muddy “C”s before. This makes the system far more robust against real-world imperfections.

In theory, AI can go even further. For instance, it could match the known elements on a plate as well as other aspects of a vehicle (such as its color, and whether it’s an SUV or a sedan) to those of vehicles which have been recently seen by the same camera. If a grey Ford F150 went through the same gate around the same time yesterday with a clean plate that reads ABC-777, then today’s F150 with a dirt-obscured first character but otherwise similar plate is very likely the same vehicle. Given enough clues like that, AI might as well stand for Artificial Intuition.

Non-AI solutions for better LPR

Optimized hardware and physical setup

Using high-resolution cameras with specialized optics can help capture clearer images, reducing the impact of distance and poor lighting. LPR suitable cameras — like Eagle Eye’s DB14 — are typically unidirectional bullet cameras, which suffer less distortion than wide-angle or omnidirectional cameras. Cameras sensitive to infrared (IR) lighting can see through some types of glare and poor visibility, as many license plates are made with retroreflective materials that are highly visible under IR light.

Placement of cameras with a relatively straight-on view of plates means less perspective distortion to deal with, and intelligently aimed light (visible or IR) gives the camera a brighter, crisper image.

Old-fashioned maintenance

For regular users, such as employees whose license plates serve as parking lot passes, regular maintenance and cleaning of the plates themselves can help the cameras do their job better. And cleaning the cameras of dust or dirt on the lens can also significantly improve system performance.

A clearer road ahead

As AI continues to evolve, expect to see LPR system’s accuracy continue to improve, in particular for cloud-based systems with improvements continually rolled in. Combining the power of intelligent algorithms with smart hardware and a few common-sense approaches means less human effort will be required just to say which vehicles are on the property.

Want to explore Eagle Eye’s AI-backed LPR solution?