[vc_row][vc_column][mk_image_slideshow images=”27389″ image_width=”1100″ image_height=”500″][mk_padding_divider size=”20″][/vc_column][/vc_row][vc_row][vc_column width=”3/4″][mk_fancy_title color=”#444444″ font_family=”none” margin_bottom=”0″]DESCRIPTION[/mk_fancy_title][vc_column_text]

How AIPARK predicts the parking

situation without a single sensor

Urbanization and growing individual mobility are globally active trends that intensify the need for transportation in cities. In this context, parking space has become a scarce resource. Drivers searching for open parking spots cause 30–40% of all the traffic in urban areas, leading to significant greenhouse gas emissions, time loss and increased fuel consumption.
Intelligent Transportation Systems focusing on the aspect of parking are a promising approach to overcome the information asymmetry and lead drivers directly to available parking spots. Thus, accurate parking information plays a key role in reducing urban traffic to tackle wasted time, noise and air pollution.

Why stationary sensors are not the golden solution

While motorists benefit from parking suggestions, traditional information systems are fully based on the installation of stationary sensors in our cities. There are plenty of sensors in the market that monitor the local parking availability in real-time. The acquired parking information is shared with drivers, typically using a smartphone app. This provenly helps drivers find the last open spot.

However, this approach has one major downturn: Installing sensors is very expensive and no one can afford to deploy such a system on a larger geographic scale.

What is a parking information system worth if it is not available everywhere? Imagine yourself trying to find a parking spot and having to handle three different navigation apps that all show sensor data for a few streets in different cities. This is about the situation with parking sensor installations today. Only equipping thousands of parking spots to cover whole city districts really creates a valuable source of parking information for drivers and eventually leads to a broad adoption of the system.

As sensors continue to be a tough investment case for parking operators and cities, there have to be other approaches that complement sensor data on a much broader geographic level. This post is meant to introduce what we came up with!

Understanding parking with traffic flow data

At AIPARK, we continuously develop new approaches to tackle the parking information problem. In order to overcome the limitations of hardware sensors, at AIPARK we are looking at new data sources to derive parking occupancy. In this context, parking-related traffic flow data or “Parking Car Data” (PCD) how we call it, is particularly interesting.

What is that and where does it come from? PCD is collected by probe vehicles that are part of the everyday traffic. We use smartphones of motorists to analyze the first and last mile of their trips to learn more about the parking situation at different locations. All collected data is fully anonymized before processing.

By aggregating PCD for target parking locations (e.g. a street or dedicated parking lot), we are able to emulate ‘virtual sensors’ without actually having stationary hardware installed in the field. To demonstrate the viability of our approach, we picked two locations in the city of Braunschweig in close proximity to one another: 1. A university parking lot (red marker), which mainly attracts daily commuters (students, faculty staff) and 2. a street with free on-street parking in a residential area slightly north of the university campus (blue marker).


red marker: university parking lot; blue marker: residential street

Looking closely at the average flows of inbound parking vehicles over the course of 24 hours the difference in parking availability at the two locations becomes visible: While peak parking time at the university parking is at about 10 am, we see increased parking traffic in the residential street during late afternoon hours. Vice versa, most residents are leaving early for work, which explains the peak for leaving vehicles at around 7 am. Students and faculty staff, on the other hand, leave the university parking lot after work or classes during the late afternoon.

Key learning: Nearby points of interest can strongly influence the time-dependent parking pressure. But of course there is much more information to be retrieved other than German students show up late for class 😉

By looking at the absolute difference in parking pressure shown in the first chart for inbound vehicles, we notice a generally much higher parking traffic in the unversity lot than the residential street. Analyzing these deltas for larger geographic areas (e.g. cities) reveals a lot of information about the differences in absolute parking pressure, especially among on-street parking areas. This is crucial when working with traffic flow data, since the share of probe vehicles compared to the entire traffic volume is typically unknown. Thus, drawing actual conclusions about parking occupancy is tricky.

Making the AI derive predictions on the parking situation

To understand the parking situation on a large geographic scale, we need to compare and ‘calibrate’ these virtual sensors for millions of streets and parking lots. This is a task too complex to do it manually and impossible to be solved with a rule-based approach. Thus, this calibration is done by a deep learning model consisting of several artificial neural networks that process PCD. Throughout the development process, we found that contextual features which reflect a parking areas’ surrounding also greatly contribute to the accuracy of the obtained prediction model. So far so good, but how do we make sure that the created model is sufficiently accurate?


Recurrent neural network

Validation and transfer learning

To validate the parking occupancy model, we conducted a dedicated data collection project over 14 months in cooperation with the local municipality in the city of Braunschweig. We installed 8 traffic cameras at 16 different locations across the city area to measure the local parking occupancy over time. Each traffic camera captured images of the respective parking area at a 5-minute-interval. The cam eras were switched to different positions once per month. To handle and analyze this large amount of data, we trained and applied an object detection model that is specialized on detecting parked vehicles.

left: Measurement points across the city of Braunschweig, right: analysis snapshot

This is how we collated a total of about 970.000 data points to validate the ‘virtual sensors’ against real local measurements. The chart below shows how the virtual measurements compare with the actual occupancy data.

What about cities with no sensors for validation?

Okay, the model seems to work well in Braunschweig, but how about other cities were we have no hardware sensors in place to calibrate and validate?

AIPARK is using a transfer learning approach, similar to the process of re-training let’s say a convolutional neural network on a new dataset with new classes: The data and the application area might change, but the underlying fundamental principles and relationships stay the same (e.g. “students at university show up late for classes” or “residents tend to leave for work early and return in the afternoon”). As long as enough PCD and contextual data is provided, we can apply the trained, calibrated and validated models also in other cities.


A quick sum up of our findings:

  • MAPE (Mean Absolute Percentage Error) across all measurement points is at 11 %


where At is the actual valud and Ft the predicted value
  • A baseline model that considers only static PCD parameters works but can in fact not stand by itself, because these parameters alone fail to explain sufficient amounts of the general variance in parking related traffic.
  • Models that consider static and dynamic contextual and pFCD data perform best.
  • Predictions in the near future (next 1–8 hours) are notably more reliable compared to predictions over longer time frames (24–72 hours).


In this article, we highlighted how traffic data reveals crucial findings on the availability of on-street parking — in terms of real-time feeds and near-future predictions. The system is highly scalable and works nationwide at dense spatial coverage without having stationary hardware sensors installed in the field. Moreover, we lined out the process of validating the parking prediction model and determining its accuracy.

With AIPARK, we can enhance all kinds of mobility services with the best predictive and real-time parking availability information. We help our customers build better solutions with dynamic last-mile guidance directly to the open parking spot. At the same time, cities can extend capabilities of todays off-street parking guidance systems by additionally including (paid & free) on-street parking, which will greatly contribute to reducing parking-caused traffic.

Realtime parking situation Berlin, April 20th 18 at 6 pm (blue icons: vacant parking spots < 5min, parking area coloring: degree of occupancy)


AIPARK is a German university spin-off startup company based in the city of Braunschweig. The companies purpose is to enhance today’s mobility services by adding best-in class predictive and realtime parking data to their offerings. Mobility services can access AIPARK predictive and realtime parking via the ParkingCloud® API . The API is currently available in 140+ German cities with prospected EU wide coverage until late 2019.

AIPARK was founded and developed by a team of six engineers, data and computer scients.[/vc_column_text][/vc_column][vc_column width=”1/4″][mk_fancy_title color=”#444444″ font_family=”none” margin_bottom=”0″]CATEGORY[/mk_fancy_title][vc_column_text]

[/vc_column_text][mk_padding_divider size=”20″][mk_fancy_title color=”#444444″ font_family=”none” margin_bottom=”0″]DATE[/mk_fancy_title][vc_column_text]
[/vc_column_text][mk_padding_divider size=”20″][/vc_column][/vc_row]
Station F Paris AIPARK