Deep Learning Is Here to Automate Parking Enforcement

A new system achieves 99 percent accuracy in identifying illegally parked vehicles.

If you live in a city and own a car, you're familiar with the calculus of illegal parking. How can we maximize our time in an illegal parking spot while minimizing the chances of getting caught, and, thus, ticketed or towed? Let's just ignore this LOADING ZONE sign, it will only be a minute. Or let's let the meter run out, it will be fine. Or maybe I didn't even see the fire hydrant.

Such formulations exist because parking rules can really only be enforced piecemeal. With a limited number of humans available to monitor parking spaces, there remains a pretty good chance that we're going to get away with parking misdeeds so long as we don't push our luck.


But this is 2017 and there is a deep learning solution to everything, or so it seems. This includes parking. Indeed, last month at the 2017 International Conference on Deep Learning Technologies a group of researchers from Xidian University in China demonstrating a deep learning-based detection system for illegal parking that achieves 99 percent accuracy.

From the deep learning side of things, this isn't terribly mind blowing stuff. In fact, much of what gets hyped in deep learning circa 2017 reduces to the same root problems, such as object recognition, speech recognition, and natural language processing. Detecting illegal parking is an object recognition problem, clearly. And like many object recognition problems, it's made more difficult by the nature of prediction-clouding complex environments, such as those imposed by weather.

Xie et al

The system described at the conference uses what's known as a Single Shot MultiBox Detector (SSD). This is a type of deep neural network that simplifies the object detection task by building up from a set of predetermined "default boxes." When a prediction is to be made, the algorithm compares the contents of those boxes to some base level. A predictive model is generated through observations of objects in relation to those default boxes.

The Xidian's group method tweaks the standard SSD in a couple of ways. The key thing is that the standard SSD is made to account for varying scales and aspect ratios of objects and default boxes. It's meant for general object detection. When detecting cars, however, things are a bit simpler because, well, cars are a lot alike. This reduces computational complexity and is what allows for the new system to operate in true real-time.

Illegal parking detection isn't a new problem by any means. In fact, a experimental program was implemented in Maryland a few years ago, achieving accuracy rates of about 95 percent. A critical difference between that system and most other methods is that it was based on magnetic sensors rather than visual object recognition. As a report on that system noted, this allows for privacy to be maintained by default.