License Plate Recognition

The License Plate Recognition system, in short, LPR, is also known as ALPR, an acronym for Automated License Plate Recognition, or ANPR, which stands for Automated Number Plate Recognition. Regardless of its name, this kind of system analyzes the image of a vehicle’s license plate, captured by a TV camera, and recognizes what numbers and letters are represented on the plate. An automatic barrier, or gate, usually connected to the system, will open, granting access to vehicles that have valid access rights.

read more

Most of these systems allow the end-users to create their own list of vehicle license plates that should be given access to specific areas, including any limitations (access only at certain hours or only into specific parking areas). Some systems also provide so-called “blacklists” for those who should not be granted access or should generate alerts for the security team.

While there are many types of LPR systems, fixed, mobile or portable, we will analyze here only fixed ones, mostly used in security and parking management.

Components of the LPR system

To capture the vehicle plate number, LPR systems typically utilize a combination of fixed TV cameras, backend software running on a local server ( Metrici, Avigilon, Gentec), or in the cloud (i.e., Avigilon Alta, Eagle Eye), and the applications that process the data, CSV, JSON, or webhooks to connect to the system for alerts or for action (like opening an automatic barrier).

Usually, when you implement an LPR system, you may need one or two TV cameras per lane (two if you have an entry-exit lane ), a server software application, a fast, heavy-duty automatic barrier, or a gate operator, and some hardware/controllers that control the barrier or gate operator.

It is also possible to have some extra hardware like Semaphores,  Parking Displays (that show real-time the available parking spaces), a voice/video communication system, and/or self-pay parking kiosks ( if it is a Paid Parking Lot), etc.

These systems  are offered in two variants:

Dedicated Turnkey LPR –  includes specialized LPR cameras, and software already installed on a dedicated server, all from a single manufacturer. They are intended for light (one two lanes) and high-traffic (e.g. multi-lane highway) applications.

Sometimes the supplier includes in his offer also heavy-duty, specially tested, fast automatic barriers.

Software-Only LPR – often integrated with other business systems, consists of specialized software (licenses) that can be used with regular, but good-quality, TV cameras (or existing IP cameras) and is usually based on deep learning algorithms. They are primarily intended for low-speed (below 50 mph) applications.

Software-only offerings have greater flexibility in camera choice but risk decreased accuracy if inappropriate cameras are chosen.

Different prices? – Not all LPRs are created equal

Many factors impact an LPR system’s price: from the camera’s quality to the speed and quality of the computing engine of the servers, which process the data as it happens. A key differentiator in LPR’s solutions is the backend software used for capturing and identifying plates. So, comparing inference speed, image processing speed, and the number of images captured per vehicle will help create a better picture of the superior system.

How does an LPR work?

While LPR systems have used optical character recognition (OCR) for more than 20 years, the recent use of Machine Learning* and Deep Learning** technologies has brought substantial improvements to these systems.

* Machine Learning (ML) ( Techtarget source) „is a type of artificial intelligence (AI) that allows software applications to become more accurate at predicting outcomes without being explicitly programmed to do so. Machine learning algorithms use historical data as input to predict new output values”.

** Deep Learning (DL) ( Techtarget source):is a type of machine learning and artificial intelligence (AI), based on artificial neural networks, that imitates the way humans gain certain types of knowledge. Deep learning models can be taught to perform classification tasks and recognize patterns in photos, text, audio, and other various data. It is also used to automate tasks that would normally need human intelligence, such as describing images or transcribing audio files”.

Let’s analyze different types of systems depending on the method used to read and process the data on the license plate:

Usually, the plate recognition process consists of two fundamental steps:

The system must first “identify” that license plate, which has the characters of interest on it, even though the TV camera can “see” many characters in its field of view (from logos, company names on vehicles, outdoor billboards, etc.).

Once the plate is found, the characters must be read, which may not be an easy task either, as it is influenced by various factors such as angles, fonts, background images, damage/dirt, etc.

Systems types

OCR-based systems recognize numbers and letters through images from high-end, expensive hardware, but with an average accuracy range of only 80-85%.  

In short, OCR-based systems take a single character as an input and either match it against complete characters or turn the character into pieces and match those pieces against patterns for characters.

OCR systems are generally quite sensitive to steep angles of view or obstructed characters. They perform best when working with plates that have a high-contrast monochrome background, large clear block letters, and evenly space between characters.

An important weakness of OCR techniques is some characters that can be confused, called visual equivalents; for instance, O vs. Q vs. 0 (zero), S vs. 5, I vs. 1, and B vs. 8, etc.

Let’s take, for example, the number „B 758 QVR”. An OCR-based LPR system could read the plate number as „8 7SB 0VR” or „B 758 OVR”, etc., which makes it more difficult to identify that license plate on a watch list.

However, the real issue is when the rate of missed license plates (ignored, unrecognized, or undetected license plates) is high, and this rate needs to be decreased. Attempting to do so may instead lead to an increase in the false positive rate, meaning that vehicles with „equivalent” numbers will be granted access to areas where they should not be.

Another drawback is that the size of the number plates and the speed of the vehicles significantly influence the resolution requirements of the TV cameras for this type of system. Between 350 and 900 PPM (pixels per meter) are required for reasonable performance, which means that for a 52 cm License Plate ( a Romanian License Plate length) attached to a stopped vehicle, there will be a minimum of 175 pixels per LP needed. So the TV camera position is very important.

Artificial Intelligence (AI) and machine learning bring a new and innovative approach to license plate recognition, leading to a higher accuracy rate of  99.02%. 

These new technologies use deep learning algorithms to read plates, however, most commercial LPR systems are hybrid systems, using a combination of machine learning, deep learning, and OCR.

There are several variants of such systems:

Deep Learning Detection – hybrid systems that use deep learning technology to detect the plates (no vehicle detection) and then use OCR to read the characters.

Deep Learning For Plates systems detect plates and read characters.

Fully Deep Learning detects vehicles, detects plates, and reads characters using a neural network trained specifically for LPR.

Deep learning is more accurate than OCR and less sensitive to variation. It deals better with smudges, plate variations, lighting changes and different angles.

Sometimes these systems provide analytics like vehicle classification based on colors and/or make and model.

OCR-based LPR uses machine learning to find plates, using heuristic cascade classifiers, just like in facial recognition, to find the features and edges of the license plate. Many machine learning detection algorithms have pre-programmed ratios for expected license plate height and width (for instance a Romanian number has 52 x 11 cm).

Machine learning-based LPRs are widely used as they offer sufficient accuracy, especialy for low-speed applications and in well-lit environments. Moreover is computationally easier than deep learning methods.

Accuracy Issues

Many LPR suppliers usually offer a „True Positive Rate” for accuracy. However, this is a misleading parameter because it ignores the number of license plates that are not detected.

For instance, a “99%” accuracy in recognizing the alphanumeric characters on a license plate means only plates that were read incorrectly count against that metric. However, what about the number of missed license plates (plates that were not seen at all), a common situation in the real world?

Note that detecting plates is generally more difficult than recognizing the characters on plates that have been detected (even than facial recognition). One reason is that license plate detection is primarily performed outside, typically in uncontrolled lighting and weather conditions, on moving subjects.

The most common aspects that cause LPR challenges are:

The angle of capture is the most critical, controllable aspect of LPR. Direct angles lead to higher accuracy compared to sharper angles. So, the TV camera’s position is important, as both vertical and horizontal angles should be as direct to target vehicles as possible

Usually for parking entrances, the „vehicle speed” is easily solved using bumpers to decrease the car speed to a level that the TV cameras will provide proper visual information. For high-speed LPR applications on highway tolls, a special TV is typically mounted directly above the road surface and not off to the side.

Damaged, bent, or dirty plates are problematic for machine learning and OCR LPR because those methods rely on finding edges. They commonly cause missed detections, missed reads, and partial reads of detected plates. Another problem in winter could be full coverage of the front number plates and even partial or full coverage of the rear plates of the cars. While deep learning algorithms are better at dealing with damaged or dirty plates because they are not looking for edges, damaged and dirty plates decrease the quality of the image captured by the camera, decreasing accuracy in the end.

Wind, rain, snow, sleet and hail, and dawn and dusk, are just a few environmental conditions that can impact the performance of LPR systems.

Low-light LPR has two major challenges: dynamic lighting from headlights/taillights, and capturing enough light to read plates. LPR cameras need to be particularly strong at WDR (Wide Dynamic Range), which keeps bright light sources darker while keeping the license plate visible. Another challenge for low light LPR is balancing camera shutter times to capture sufficient light, but not so long that the subject suffers from blurring. Using IR TV cameras with filters for visible light could be a good option.

License plates vary widely throughout the world, with differing sizes, lettering, coloring, and images/backgrounds. Europe plates are more uniform and, in general, follow common formats (generally white with black lettering); therefore, they are much easier to read; however, even in these cases, there could be variants that challenge the LPR systems.

When considering an LPR  provider, it is important to check for support for your geographic region. This is typically on a country-by-country basis. However, even within a certain country supported, if certain cars within that country have different types of plates, accuracy could be significantly reduced.

Benefits Of LPR Data

Most complex LPR solutions today capture more than license plate data. They can also identify the vehicle’s make, model, and color associated with the license plate. But like all data, it is only valuable if you use it.

Some of the most significant uses today include facility management and perimeter security, automatic management of gates, entry/exit, management of paid parking, etc.

contact us