By Chris Scheppmann

When we visualize artificial intelligence (AI), we often think of robots learning how to think, so they can perform human tasks. And of course, those of us who are science fiction fans probably envision apocalyptic acts committed by hordes of out-of-control robots. Thankfully, the reality is much safer and more useful than robots learning how to kick a soccer ball. This is particularly true when it comes to parking.

One of the most important recent breakthroughs in parking guidance technology is Machine Learning. Through Machine Learning, parking guidance has become highly accurate and useful, both for helping manage parking inventory and when it comes to providing parking operations with business intelligence to make better informed decisions. But, to understand the role that Machine Learning is playing in parking guidance, it is first necessary to understanding what Machine Learning is.

What is Machine Learning?

Machine Learning is a type of AI. Equipment paired with Machine Learning is able to modify itself when exposed to more data. It is dynamic and does not require human programmers or designers to manually make changes and the Machine Learning models can continually  improve its understanding of an environment where it is being used. 

As Arthur Samuel, a pioneer of the field stated in 1959, Machine Learning “gives computers the ability to learn without being explicitly programmed.” As an example, Machine Learning is like a child who is born without having any knowledge and adjusts (knowledge improves) its understanding of the world in response to experience (receives new data). As that baby continues to be exposed to similar and new experiences, its ability to make connections and decisions improves. Over time, a child can differentiate between a spotted dog and a cow or a brown-haired dog and a horse. 

Machine Learning uses algorithms that predict certain outcomes and minimize errors in those predictions. They rely on neural networks, to keep measuring errors and modifying the program to minimize—and hopefully eliminate—those errors. These neural networks, which are modeled loosely on the human brain, allow the technology to learn to perform a task by analyzing training examples. An object recognition system, for instance, might be fed thousands of labeled images of cars, trucks, motorcycles, buses, etc., so it can find visual patterns in the images that consistently correlate with particular labels. In short, it learns how to differentiate between objects by analyzing the shapes of different objects.

Machine Learning in Parking Guidance

That is exactly how Machine Learning improves parking guidance systems. The technology is programmed into parking guidance equipment, not just to process the video frames, but also to make decisions based on the frames. This is done with a neural network. 

This type of approach typically demands extra and specialized processing resources and because of that, Machine Learning has only started to become a possibility in parking guidance in the last few years. With Machine Learning, you typically provide the neural network with many sample images to train the network. The system then “learns” to recognize the trained objects in images. This is a much more accurate and precise approach when you need detailed information about what is in an image. For example, a Machine Learning system that is trained to recognize vehicles, will take an image, and look for all instances of the items it was trained to look for and then let the system know how many of those objects it found and where they are in the image. It can even provide a confidence score associated with the objects. 

When it comes to identifying different vehicles, the neural network will learn the dominant identifying features of various kinds of vehicles, such as cars, motorcycles and scooters, and delivery trucks and look for those features as cars enter the parkade and travel through individual sections of the facility.

An additional benefit to Machine Learning is the confidence score. If an object is detected but has a low score, it is still detected. While the guidance may not include it as an object of interest initially, it can later evaluate all the objects with a low score to see why the score was low. Taking advantage of this allows the system to recover from low scores and continually train the network’s intelligence over time. It is an important element that allows the system to continually improve itself.

Why Machine Learning Matters

The primary advantage of machine learning is accuracy. The benefits of continuously improving a system’s accuracy are obvious. The more accurate a guidance system is at reading what type of vehicles are in the parking facility, the better able the system is to offer accurate occupancy information. 

How big of a difference does it make? The traditional benchmark for single space guidance systems has traditionally been 99%. Single space guidance providers have relied on this increased accuracy to entice owners and institutions into spending upwards of hundreds of thousands of dollars for single space systems. With the emergence of machine learning, intelligent cameras that perform zone and level counting have reshaped the reliability and accuracy expected and achieved with these systems. The watermark traditionally has always been 99% per counting node, which few systems actually ever achieved. Now with some purpose-built vision-based technologies, each node can consistently achieve mid to high 99%. This is possible through continuous learning with machine learning.

That is particularly important at a time when parking lots and parkades are being expected to provide service in addition to parking vehicles. For instance, many parkades and lots offer space for delivery vehicles. At a time when curb management is a key to mobility, a parking guidance system’s ability to recognize distinct types of vehicles is essential, both for quantifying occupancy and space availability and better managing the curb.

Improving accuracy through Machine Learning also offers significant cost benefits to parking owners and operators. In the past, owners and operators needed to install single space guidance technology to enjoy sufficient accuracy. The problem was it could cost anywhere from $500 to $1,000 per space for a single space system. So, for a parkade with just five hundred parking spaces, the cost of installing a reliable guidance system would run around $250,000 and could run as high as $500,000. And that is just the initial installation cost. There’s also maintenance, service, and other costs to factor in.

This is particularly important for owners who are considering retrofitting parking facilities with parking guidance technology. Owners can now enjoy extraordinary parking guidance accuracy without having to install expensive and intrusive infrastructure. With a camera-based system, installation and on-going costs are reduced significantly while providing the same accuracy levels associated with single space systems. And, since a camera-based system is continuously learning, it can easily detect parking violators where a single space system will only tell you if the spot is occupied or not. As a result, accurate guidance and valuable asset utilization data can be achieved for a fraction of the cost of traditional systems. 

The Future of Parking Guidance

Machine Learning represents the future of parking guidance. And when it comes to parking and AI, the future is not nearly as scary as science fiction authors would have you believe. In fact, through the AI elements of Machine Learning, parking guidance has become much more efficient and cost-effective, and can play a much more significant role in promoting mobility through curb management.

About the author:
Chris Scheppmann is managing member of EnSight Technologies, a leading parking guidance provider. He can be reached at chris@ensight-technologies.com.

SHARE IT:

Leave a Reply

You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>