Separating Fact from Fiction: An Insider’s Insight into LPR

Separating Fact from Fiction: An Insider’s Insight into LPR

By Bill Franklin

Our company, Tannery Creek Systems, has been making License Plate Recognition equipment since 2007. In the sixteen years growing our company across North America, one of the most common and interesting topics that our clients raise is license plate recognition (LPR). How does it work? Can I use it for permits? Do you sell this data? What is the rate of recognition?

This article was inspired by the LPR panel that I moderated at this year’s Canadian Parking Association conference and tradeshow. The set of questions raised by or to the audience is the springboard to this article. I have added additional questions that our clients have asked us. In this question-and-answer format, you may discover a deeper understanding of LPR.

What is LPR?

License Plate Recognition (LPR) is a specialized Optical Character Recognition (OCR) system. It requires a camera and computer to extract the characters and try to make sense of them. LPR, among other names, is also known as:

  • Automatic Number Plate Recognition (ANPR) 
  • Automatic License Plate Recognition (ALPR)
  • Automatic Vehicle Identification (AVI)
  • Plate scanner

LPR was originally developed at Scotland Yard in the UK and began operation in 1978, so it has been around for a good many years.

How does it work?

Typically, a camera captures 20 frames per second (or more) and these frames are analyzed for the presence of a rectangular surface with digits, i.e., a license plate. The analysis can be in the camera that has a sophisticated microprocessor, a graphics processing unit (GPU)  or processed in the Cloud which will employ essentially the same thing, i.e., a GPU. Two competing systems perform LPR: font-based recognition versus the newer artificial intelligence (AI) also called machine learning based systems. 

Originally, the LPR process compared each character captured by the camera to digits in its software library. For example, an A will be compared (correlated) in software to the LPR’s internally stored library and the letter A. Since fonts vary, one can see that this process can be a little complicated as the following example shows:

Characters match:

Captured plate   ABC123

Compared to    ABC123

But how about this example? The font based matching system will not match.

Captured plate ABC123

Compared to  ABC123

Font based LPR schemes store North America’s various license plate fonts and compare the letters in the order that is most likely. For example, in Ontario, the license plate for Ontario is compared first. Each jurisdiction will tweak their fonts for aesthetics or other reasons. Font based LPR systems have serious weaknesses and struggle with blooming, shadows, low contrast, and other photographic artifacts.

Artificial intelligence (AI) LPR systems are superseding font-based systems. With AI, one can “teach” the AI system (also called machine learning) so that ABC123 does indeed equal ABC123. AI systems are typically trained (i.e., taught) with hundreds of thousands of samples. Generally, the greater the number of samples, the more accurate the system.

A peculiarity of AI is if we teach the AI that the captured plate ABC123 is actually ABC133, then the AI will return ABC133 each and every time it is presented with a plate that is ABC123. Hence it is vitally important to do good quality training of machine learning. You can also reach out to your LPR vendor to have them “reteach” for an errored plate.

How many cars can you scan in a day?

Many vendors state 20 frames per second or even 80 frames per second (implying hundreds of thousands of vehicles per day), but the truth is more complicated.

For example, our products can scan and analyze from two vehicles per second to about 20 vehicles per second. But driving at 40 km/hour means at most you can scan two parallel parked vehicles per second on the right side of the patrol vehicle and two on the left side if the road is not too wide and the view of the vehicles is not blocked. Hence four vehicles per second is pretty much the limit.

What happens when the patrol vehicles get ensnared in rush hour traffic? Or on a crowded beach parking lot. In this case one is lucky to drive faster than 2 km/hour! 

In a parking study in a small city near Toronto we analyzed 35,000 parallel parked vehicles. The average time per vehicle was 10 seconds because of traffic, pedestrians, signal lights, etc. At 10 seconds per car, one can scan about 360 vehicles an hour or a little less than 3000 a day. In practice with lunch breaks, and interruptions, 2000 vehicles per day is a good estimate. 

Parking lots will be roughly two to three times this figure if congestion is not too bad and the parking lots are in close vicinity, i.e., 4000 to 6000 vehicles a day. Again, it depends. Small parking lots with blind runs are less efficient to scan. Hence interspersed small parking lots may be no better than 2000 vehicles per day.

Finally ticketing takes time. Issuing 30 citations can take 60 minutes. In that case the number of cars scanned in a day is reduced to 7 hours of scanning instead of 8. This can be improved by having walkers issue citations, but the labour is the same when one includes the walker’s time issuing citations.

Do you need two people or just one in the LPR vehicle?

Most systems only require one.

What are hot lists?

Hot lists are lists of license plates of vehicles that are wanted by police or for scofflaws or other offenses that are deemed serious. Typically, LPR systems can handle many hot lists. These lists are downloaded automatically over the broadband internet and into the mobile LPR computer memory.

How many lists can you have and what type?

Other lists include permit lists such as Residential Parking Permits, monthly subscription lists, VIP, handicap, contractors. Pay by Plate can also be considered a type of list. All of these are downloaded or updated automatically. Updates to the lists are done periodically and are called delta updates, i.e., only changes are sent to reduce the volume and time of data transmission.

What is a good LPR read rate?

Typical competitive read rates are 95% or better of eye readable plates Dirt, snow, glare and smoked plastic license plate covers tend to reduce the rate.

What is the difference between detection rate and LPR read rate?

When testing the yield of an LPR system, it is useful to know how many vehicles are scanned. This can be significantly more than the LPR system reads, for example during the winter or where license plates are only on the backs of vehicles. So, the real yield of LPR is the number of plates accurately read divided by the total number of vehicles it has passed, whether or not the LPR system detects them. If eye readable plates are 95% of the vehicles scanned, and the LPR is perfect at reading eye readable plates, then the LPR yield is actually 95%  not 100%.

Can LPR read through snow?

We have been asked this question many times. The short answer is, if you cannot see the plate by eye, the LPR system will not see it either. Most LPR systems use Near Infrared or natural light to detect plates by detecting a rectangular flat surface with alphanumeric digits. If it is covered with snow the LPR system will not see this vehicle. Tannery’s top end equipment uses a laser to detect vehicles, but the driver still must manually enter the plate or brush off the snow and drive by scanning the vehicle again.

Plates that have, let’s say, 10 to 20% snow coverage are typically read, but it greatly depends on the vendor and the location of the snow on the plate.

Can LPR detect vehicles without plates?

Most LPR systems cannot detect vehicles without plates. In fact, in our experience, about 2% of vehicles on average do not have legible plates. This increases markedly in the winter where we have seen 20% of vehicles without legible plates. Where license plates are only mounted on the rear of vehicles, drivers will begin to back into parking stalls dramatically increasing the number of vehicles without visible legible plates.

Will covers on plates affect the read rate?

Yes, if the cover is tinted and particularly if the cover has salt or dirt. It all depends on the legibility of the digits visible to the camera. We see roughly a 10% drop in yield for grey smoked plastic covers and a steeper drop in the winter.

Does dirt or salt affect the read rate?

It depends on the quantity of dirt. If the digits are not legible by eye, then the yields will be low. If small parts of the digits are covered the LPR rate will still yield 80% or better. Salt can have a significant impact on LPR read rates for vendors that exclusively use near infrared.

Does glare from the sun affect read rates?

Glare from the sun reflecting off the plates can impact LPR because the plate looks totally white. Generally, the impact of the sun when averaged over the day is relatively small.

Can I use my LPR system to do parking studies?

LPR is highly effective for parking studies where plates are present on the front and back of vehicles. It is less effective where only a rear license plate is visible. Tannery provides a parking study application, and some cities extract the data from autoChalk’s Amazon Web Services database (each client gets its own database and extracts the data to perform their own parking analysis).

How do privacy laws affect LPR?

This interesting question on an acutely sensitive subject depends on the vendor and state or provincial legislation. In general, we have found cities tend to be circumspect on retaining the data or sharing it and especially selling it. But it depends on the jurisdiction and the local sensitivity of citizens to the LPR scans.

Is a city’s LPR data sent to other parties
e.g., private companies like private investigators or other governments?

Some vendors will sell the data. We recommend cities think carefully about the LPR data and the implications of sharing or selling the data. Tannery does not share or sell data as we view it as the city’s property.

How do people game the system to avoid being caught with LPR?

Parkers routinely cover their plates, use dark covers, let the paint peel, install (unused) bike racks or carriers on the back of their vehicle, apply reflective paint over the entire plate or alternatively light absorbing paint (this is typically ineffective for LPR that uses natural light).

What factors affect reliability of a parking system that includes LPR?

LPR  systems are generally part of a bigger system. The reliability of the system depends on the individual subsystems and the overall interoperability of the subsystems. Take Pay by Plate for example. It will often include:

  • Pay stations
  • Pay by phone or app
  • Payment clearing house
  • Broadband carriers (e.g., Telus, Bell, Rogers)
  • LPR
  • City information technology (IT)
  • The users (parkers)
  • Parking enforcement and more generally management
  • Decisions by the city (including elected officials) in terms of policy and signage which can make the system simpler or more complex
  • Local businesses and business membership groups
  • Organizational aspects for instance help lines, vendor error reporting and response and so on
  • Moreover, parking is a favourite target of the media and politicians scoring points. Getting it right is important, getting it wrong can be a public relations disaster.

 A good system will excel in simplicity, reliability, and responsiveness, and make all stakeholders proud. A good LPR partner is excellent at integrating and a good team player. These attributes help to solve faults and failures or, even better, avoid problems from the start.

Does LPR make my staff more productive?

An efficient and reliable LPR system will make your staff significantly more productive compared to manual chalking or walking. Cities typically achieve 2 to 3 times more productivity. Productivity is highest when the streets are not congested, but overall, vehicle mounted LPR can scan faster and is more resilient to inclement weather compared to walking PEO’s. 

Does LPR make my parking enforcement staff impersonal and remote to the public? 

PEO’s cocooned in a patrol vehicle will seem more distant and remote. On the other hand, confrontations are less frequent. A city can make their parking management more approachable by having some officers on the beat. However, when scanning Pay by Plate sessions, vehicle mounted LPR is hard to beat.

Can an LPR system apply different parking zone policies automatically? 

LPR systems often have GPS awareness and ideally can adjust to different parking zones and their respective policies. As the LPR vehicle drives through the zone it adjusts its parking policy on the fly, for example changing from Zone A’s 2-hour parking duration to Zone B’s 20-minute loading zone policy. Automatic and adaptive zone awareness is very desirable as it makes the PEOs more efficient and less likely to make an error.

How does LPR impact appeals?

When equipped with megapixel photos, LPR is invaluable in administrative appeals and court action. Detailed background imagery provides context and makes the infraction evidence more compelling, discourages frivolous court cases and is particularly useful for digital chalking. 

What does AI mean and how does it affect my LPR?

LPR is increasingly using artificial intelligence (AI) to recognize plates and even vehicle types and presence of vehicles. It is more resilient to partial shadows, low contrast, dirt, characters that are clipped by a vanity frame and other aspects that make plates hard to read. AI also teases out parking patterns offering insight as to where to focus enforcement, increase parking availability or where parking fees can be adjusted to reflect demand. For example, autoChalk uses AI based LPR to achieve high recognition rates.

Is LPR effective in provinces where the license plate is only on the back of the vehicle?

LPR is less effective in jurisdictions with plates only on the back of vehicles. As soon as people realize LPR is at work, they change their behaviour to thwart the detection of their license plate. Tannery’s top end equipment uses a laser to detect vehicles without plates, but the officer will still need to get out and check the plate.

Does LPR work with non-reflective plates?

LPR that uses white light (i.e., natural light) is unaffected by non-reflective plates. Near Infrared (NIR) systems are generally seriously affected by non-reflective plates and typically do not read or detect a non-reflective plate.

How does very cold weather affect the equipment? Also, other weather conditions?

LPR equipment can malfunction in extreme cold (-30 C and colder). The glass lenses can get coated with frost or ice and the pictures are affected. Heavy rain, snow and occasionally dense fog can affect reads and yields.

Can I apply LPR to these categories?

a. Pay by plate

b. Permits

c. Handicapped

d. Residential parking permits

e. Street VIP permits

f. Parking lot permits

g. Garage permits

h. Street storage

i. Digital chalking

j. Scofflaw

k. Stolen vehicles

LPR can be used for all these items. Check with your LPR supplier.

What is fuzzy matching?

Fuzzy matching is a term used to describe matching on plates that are almost the same but differ typically in a couple of digits. For example, people that suffer from dyslexia will flip two digits, like this:

Actual plate = ABC123

Entered plate = ABC213 

The LPR system looks through the list of paid license plates. It does not find ABC123 but does find ABC213. A fuzzy logic equipped LPR will then treat ABC213 as if it was ABC123 and treat it as paid. This creates complexities, e.g., when both plates are scanned, i.e., ABC123 and ABC213 are parked but only one is paid, say ABC123. One can see the case where two people have dyslexia and each have switched their digits. Who paid?

In any case, some of our clients allow fuzzy logic and others do not. 

Is LPR effective for secure admission into a garage?

LPR is not secure at all. It is easy to spoof a license plate with widely available commercial tools such as colour printers, scanners, photoshopping software and such. 

However, if the goal is simple admittance to a garage and the risk of inadmissible people entering is low or their impact is not significant, then LPR is a good tool for admittance.

If security must be stronger, then RFID is a better choice as it is harder to fabricate a device with the correct protocol and serial number and spoof the RFID system.

Finally, snow, salt and plate covers may affect the plate recognition hence a keypad entry with the user’s password will still be required.

As you can see, license plate recognition is a powerful tool. Understanding its strengths and limitations can help you get the most out of your LPR investment or invest in the right LPR equipment that suits your needs.


About the author:
Bill Franklin, P.Eng. President, Tannery Creek Systems

since 1983 in the fields of computer and electrical engineering, data communications, information security and parking automation. A registered professional engineer, Bill managed product development groups and business units in the role of Vice President before starting Tannery Creek Systems. His clients have included RIM (Blackberry), XM Satellite-Sirius, Honeywell, IBM, Bell Telephone among many others.



Leave a Reply