AI AND VIDEO ANALYTICS BLOG
Video Surveillance & Physical Security Industry Viewpoints
August 26th, 2020
Author: Yogev Wallach

A Buyer's Guide to “In the Wild” License Plate Recognition

What is License Plate Recognition “In the Wild”?

License plate recognition has become an important part of everyday operations in fields like law enforcement, corporate security, and highway management. Various agencies and corporate security teams need to identify vehicle license plates for a number of reasons, from collecting tolls, conducting post-incident investigations, or tracking license plate numbers and vehicles that commit traffic violations. By leveraging “In the Wild” license plate recognition capabilities as a fully integrated component of a video content analytics platform, organizations can utilize existing CCTV networks in place along highways, streetlamps, or buildings to capture images of vehicles, and identify license plates based on configured watchlists or other video footage to drive some of the above-mentioned use cases and more.

Powered by deep learning and artificial intelligence, video content analysis enables complex object extraction, recognition, classification, and indexing activities, thus making video searchable, actionable and quantifiable. Operators can conduct fast, filtered searches of video footage across multiple cameras, based on vehicle characteristics, direction, and license plate recognition technology (LPR) also known as “automatic number plate recognition” (ANPR). This post will explain how LPR can be used to enhance intelligent video surveillance by explaining the differences between traditional, constrained LPR scenarios and solutions versus LPR “In the Wild.”

Constrained vs. “In the Wild” scenarios

There are two basic categories of video technology solutions for LPR: 1) constrained, and 2) “in the wild.”  In constrained license plate recognition scenarios, the following factors are at play:

  • A specialized LPR camera is usually used (sometimes with additional active illumination).
  • The camera has a single purpose and is positioned in a dedicated way (e.g. so the license plate appears large in the frame of the camera).
  • The camera is in a controlled (constrained) environment where one can predict how the vehicles will behave, such as travelling directly towards the camera – or even slowing or stopping at a barrier in the entrance or exit of a parking lot.

In constrained scenarios, the license plate recognition accuracy rate is usually high, because the camera is ideally positioned, usually at the vehicle’s level, with its field of view calibrated towards the expected location of the plate. Furthermore, the lighting is optimized, and the camera located in an area where vehicles typically stop – as is the case in a parking garage exit, for example – so there is no motion blur. Another example is a toll road billing system where vehicles can be captured in full speed – but the profile is known in advance, so the camera positioning and setup, and lighting are designed accordingly.

In contrast, “in the wild” license plate recognition detects and transcribes license plates using generalized video surveillance infrastructure and environments. The camera selection, position/placement, and setup are not dedicated to LPR functionalities, which make it more challenging for the cameras to detect and transcribe license plates. The lighting is not always controlled and optimal, and the field of view is often wide, which means the image contains far fewer pixels per plate. Lastly, in these scenarios, the quality of recorded surveillance video is usually optimized for storage efficiency rather than for quality which make it more challenging to extract information.

Factors to consider when choosing LPR technology

“In the wild” scenarios are much more challenging for LPR because of these limitations listed above, as well as everyday situations, including:

  • Obstructed and obscured plates caused by scenarios, such as environmental conditions – i.e. mud or snow may obscuring some of the letters or numbers on a license plate – the frames around the plates covering some of the text on the plate, or plastic covers on a plate creating glares.
  • Different linguistic characters and fonts on plates in different states and countries
  • Discerning between similar characters such as a zero (0) and an O, or 8 and B.

Not all of today’s intelligent video surveillance systems are designed to overcome all of these challenges and provide accurate, robust LPR systems, based on standard video surveillance cameras.  Therefore it is important to compare the operational envelope when comparing accuracies and prices.

Some LPR solutions today come bundled as part of a suite of analytics.  Alternatively, there are stand-alone LPR technologies that offer limited functionality, such as simple access control solutions that serve small parking lots. Organizations will certainly get more return on their investment by choosing a comprehensive and extensible video content analysis solution that includes LPR, because, in so choosing, organizations can leverage many additional video analytic capabilities and features that accelerate investigations, improve security and situational awareness, as well as streamline operations with business intelligence. 

Accelerating post-incident investigations with LPR

When an investigator needs to identify a specific car that drove away from a crime scene or accident, video analytics technology can be used to help police dispatch issue a “be on the lookout” (BOLO) report of a suspicious vehicle: Besides the obvious make/model/color of the vehicle, officers can include a license plate number, to assist officers in finding the specific vehicle of interest.

Using the search and filter capabilities of a video content analytics system, officers can search video footage from multiple cameras, for vehicles with a license plate that matches an eyewitness testimony or other intelligence (that provide a full or even partial plate number) and track the suspect’s activities in the days leading up to the incident or in the time after the incident. Filtering video for the identified vehicle can help investigators gather more evidence about the car and, in some cases, the driver of the vehicle.

LPR alerting for increased situational awareness

Similarly, the video intelligence system operators can configure real-time alerts whenever the license plate of interest is detected. They can also add the plate to a pre-defined watchlist and trigger real-time alerts for any plate on the watchlist detected in a video feed. Alerting logic can be set across multiple cameras, simultaneously, to help officers improve their situational awareness and response times when vehicles of interest are detected in an environment.

In addition, a user can set up a list of allowed vehicles and set a rule to be alerted when an unrecognized vehicle enters the premises.

Streamlining operational efficiency with LPR

Video content analysis can also use LPR to collect data about vehicles that have traveled between two points. Whether organizations such as departments of transportation and state law enforcement agencies want to understand whether dispatched buses, taxis, or trucks have returned to a depot after a journey (and measure the duration of the excursion), or to detect speeding trends on a toll road or stretch of highway, they can generate visual reports and research vehicle behavior patterns to quantify activity, make intelligent decisions, and improve traffic and asset management. Using regular surveillance networks, the video content analysis system can gather data by tracking the license plates of various cars as they pass by different cameras on the highway and measures the amount of time it takes to travel a certain length of the highway.

To summarize, as with all solutions, when security, transportation or enforcement teams are comparing license plate recognition technologies, it is important to consider solutions that offer consistent, accurate performance – whether in constrained or unconstrained environments.

It is important to benchmark the systems in the real-world conditions they will encounter in the field and take into account the need for a “man in the loop” (as some systems, e.g. those used for automatic parking management, are set up to require minimum human interaction – while others that serve investigators have a much better operational envelope but might require human interaction).

However, beyond those LPR benchmarks, platforms that offer other complementary functionalities, such as long-term aggregated data for improved operational efficiency, on-demand processing of stored video, and real-time alerts to improve situational awareness, enable organizations to derive additional value from their existing investment in video surveillance networks in various ways, by combining LPR with comprehensive video analytics.