About 15 years ago, I attended my first parking show working for a license plate recognition (LPR) company.
At our booth, we had our new camera, roughly the size of a small pizza box, connected to a PC and monitor. When an attendee inquired about our technology, it was showtime. I would grab a California license plate sourced from eBay, walk out a few feet from the booth to a predetermined spot in the aisle, and wave the plate as I had rehearsed several times until I heard a “beep.”
That sound indicated that the system had read the plate. Then came the moment of truth — had it read the plate correctly? I would walk back to the booth and look at the monitor together with the attendee, fingers crossed that we had read it right. In most cases (though not all), the system would read it correctly, dramatically displaying the black art of LPR magic for all to see! Back then, LPR technology was in its infancy, particularly in the parking industry, but it worked well enough to be effective for several applications.
Since then, the technology has advanced significantly, improving in terms of performance and consistency. Innovation in the LPR arena has resulted mainly from three key advancements: the use of artificial intelligence (AI), major improvements in camera sensor resolution, and significantly better lighting. Thanks to these enhancements, LPR is more reliable, and therefore more viable, than ever before.
Incorporating AI
The single most important factor in the improvement of LPR is the “engine,” a term often used to describe the algorithms performed by a computer to locate and interpret license plates. Previously, this multi-step process often involved finding a plate and then segmenting and identifying the individual characters.
For the sake of this article, we will refer to that technology as “rule-based optical character recognition” (OCR). In oversimplified terms, developers set rules that define what a letter “A” should look like, as well as “B,” “3,” and so on. Intensive development of this approach by LPR manufacturers led to major improvements. Although the performance of such systems would vary based on factors such as lighting, weather, and jurisdiction, overall, LPR still did an excellent job in most cases.
Five to 10 years ago, a premium rule-based OCR solution would achieve accurate readings at rates of between 92% and 97%. Manufacturers later included software tools — for example, fuzzy matching, which helps match plates to a predefined list of vehicle plates, such as registered parkers, scofflaws, and the like — that would artificially increase this rate to between 97% and 99%.
Although the solution was not perfect, it had a transformative effect on the parking industry. Whether enabling automated parking enforcement, reducing congestion at gate entrances and exits, or facilitating implementation of a frictionless free-flow parking system, LPR technology had an immediate, measurable impact on efficiency.
However, headaches ensued on occasions when performance fell short, the wrong citation was issued, or the VIP was denied entrance. Naturally, if you worked for an agency or operator that relied on LPR as part of its day-to-day operations, performance was critical.
This is where AI proved itself a game-changer. Unlike rule-based OCR, modern AI engines use a neural network to engage in machine learning. In this process, AI engines are fed thousands upon thousands of plate images in a technique known as “training.” After ingesting all these images, the trained AI engine becomes highly effective at identifying and interpreting plates, a process known as “inference.”
Read rates of modern AI-based LPR systems can easily hover around 97% to 98%, even without the use of artificial matching tools. The result is fewer false infractions, more gates vending automatically, more effective gateless systems, and fewer headaches.
Yet, AI has additional advantages. AI-based LPR engines can better accommodate such difficult conditions as bad weather, partially obstructed plates, and poor lighting. In addition, this type of solution adapts better to changes in license plate design than does a rule-based OCR system.
Essentially, AI engines have an easier time accommodating new types of license plates. For example, if Florida decided to introduce a new Tampa Bay Buccaneers Super Bowl LIX specialty plate next year, the effort to add such a plate to an LPR engine would vary greatly depending on the type of engine. Updating a rule-based OCR engine would likely require research and development to study the plate, identify its unique aspects, and modify and add rules to recognize and interpret the plate in the future. This process would likely require some iterations to perfect.
With the AI engine, by contrast, sample images of the plates would be included in the training process, greatly reducing the development effort.
All these features render modern AI engines better performing, more robust, and more adaptable moving forward.
A Revolution in Resolution
The second big change that has affected the LPR world is sensor resolution. Digital camera sensors have gone through a true revolution in the past 20 years. Initially a novelty, they were not taken seriously by professionals, as they had low resolutions and performed poorly in low light. Modern sensors and cameras have addressed those shortcomings and can produce great levels of detail with proper illumination. These changes have improved LPR for parking in two ways.
The first is what many refer to as the field of view (FOV), which is what the camera “sees.” Setting the proper FOV, which is closely tied to the “zoom” of the camera, entails striking the right balance between plate capture and resolution. Intuition would tell us to zoom out so that we never miss a vehicle or plate, but the limited resolution of the sensor would make for a poor plate image. On the other hand, if we zoom in too much, we get a clear plate image but are more likely to miss a vehicle or plate. Hence the need to compromise between capture performance and resolution.
However, as the resolution of new sensors increases, we can further decrease the zoom — increasing the plate capture rate — while still having enough resolution to read the plate. A decade ago, most LPR cameras had a resolution of less than 1 megapixel (MP). Today, many manufacturers offer cameras having double the resolution, at 2 MP. Although exceeding that resolution would be even better, doing so typically introduces processing bandwidth limitations.
Improving Illumination
The last factor that has significantly affected LPR is the drastic improvement of illumination, or lighting. This improvement results from a combination of camera sensors working much better with less light and light-emitting diode (LED) technology producing more light for a single LED package.
There are three main approaches to illumination and LPR. Some manufacturers use infrared light with special filters, others use white light, and some providers offer both types of light. Each approach has its pros and cons, and some premium installations utilize both side by side.
Regardless of the approach used, adequate illumination remains a critical factor. Many newer solutions that are self-illuminated can produce much more light at night than older models. The added light produces a much clearer image, resulting in more reliable plate read performance.
Ultimately, these three changes to LPR have significantly improved the technology in terms of reliability and consistent performance. This enhanced performance can contribute to improved parking operations, even compared to relatively recent solutions. Whether you are shopping for a new LPR system, or have had an existing one for a while, you should expect more today than you did just a few years ago.
Chris Yigit is a Senior Product Manager for Flash Parking. He can be reached at chris.yigit@flashparking.com.
