How does machine learning help parking guidance systems adapt to new vehicle types and new parking behaviors?
The parking world is dynamic. Parking behaviors, vehicle styles and shapes, variable in-space parking types (i.e EV charging, disabled access, compact vehicle, carpool vehicle, package delivery, food delivery services, etc.), variable pricing models, reserved and VIP access, all combine to add complexity to any operation.
But, in one place, most of these elements (and more) come together in a hyper-dynamic parking zone we call “curbside.” Initially applied to areas literally on a busy urban curb, curbside can more broadly apply to any high volume, multi-use zone. It may be managed by a public entity, as in the instance of on-street municipal parking or by a private property owner in the case of high-volume shopping, event, and multi-use areas. Because of their high desirability, curbside scenarios often experience a high degree of irregular parking events, high volume, and very dynamic user behavior.
To be useful in this environment, parking systems must have the capacity to handle a multitude of parking variables and to distinguish discreet parking events beyond simple occupancy. And many do, but often managing that process is manual and time intensive. A few, however, have not only the capacity for flexibility, but also the capability of improving their own performance over time, and with little human intervention. They do so by learning to be better, through a process called machine learning (ML).
The basic principle of machine learning is to create a system that allows a computer to discover how to perform a task without explicitly being programmed to do so. Here’s more about what machine learning is, but one simple way to think of it is as computer scientist Tom M. Mitchell of Carnegie Melon University put it: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.”
The key phrase here is “…improves with experience.” Multiple strategies can be employed to train a machine to learn, from building complete algorithms to designing methods that facilitate the device building its own as is necessary when the number of possible variables is very large. No matter the approach, as the machine learns its tasks and encounters more instances/data points, its task outcomes improve. For example, it may first learn to distinguish between a passenger vehicle, a delivery truck, and a shuttle bus. Then, over time it may be possible to train the system to distinguish between a food truck and a delivery truck. Or a taxi and a ridesharing vehicle. It is this capability that makes applying ML to curb management scenarios very effective.
ML enabled guidance systems can be trained on what defines a parking event, correlate that from multiple data points, and evolve an effective analysis to give accurate occupancy data. It can also incorporate time into its evaluation for reservations, loading zones, event parking, or other time sensitive scenarios.
ML can be applied to any data input, but for parking guidance, video analytics offers the most amount of data from the least amount of hardware and minimally invasive installation needs. While line of sight is essential for video, in most curbside parking scenarios this is easily accommodated. By using video analytics, sensors can be trained to identify different types of vehicles, each of which might have their own level of access to the curb. Delivery trucks may have a permit for a specific time of day, buses, rideshare, and taxis another, scooters, and bicycles yet another. The ability to learn and distinguish these different modes of transportation allows a high degree of automated management capability (permit verification, auto-pay events, violations) and consequently more options and capability for managing the curb than other sensor types, often with fewer pieces of physical hardware.
Additionally, as ML continues to advance, systems will be able to advance with it, at least for several generations, extending the life expectancy of the system and driving down the total cost of ownership. Plus, devices that have gained experience share that experience throughout the system, improving the system while retaining memory transfer capability and redundancy in the event of equipment failure or replacement. This means operators can rest assured that their state-of-the-art investment continues to perform in the long term.