VLOED: Can we predict city crowds?
How can we create an accurate picture of crowd levels in the city?
How many visitors and cyclists are there in the city center? What method of crowd measurement is most relevant? And where should we focus our efforts? In 2021, 2022, and 2023, Ghent experimented with crowd data through so-called living labs. These allow us to quickly try out various approaches with short-term tests.
Prior to 2021, the City of Ghent already conducted crowd measurements. For instance, during the Gentse Feesten, we experimented with measuring the festival crowds in collaboration with Ghent University. We have previously published various crowd data, and since 2020, Covid-19 has made everyone familiar with the concept of crowd levels and crowd barometers. Through the VLOED project (Predicting Local Business and Economy through Crowds), we are building on our existing expertise and exploring potential new steps. This project is supported by the European Regional Development Fund (ERDF) and Flanders, and we are collaborating with the city of Bruges.
In VLOED, we use living labs to set up simple and temporary tests to assess what is possible and feasible. This allows us to explore a range of possibilities with a straightforward setup and quickly evaluate these options.
programme
-
How are we exploring the possibilities?
-
-
What do merchants and the city want to know?
What do we need to know, and more importantly, how do we use this information?
In recent years, Ghent has gained significant experience in crowd management through experiments with crowd measurements during major events such as the Gentse Feesten and the Light Festival. At the start of the VLOED project, we reassess this experience. What exactly does the city want? What is useful for merchants? And most importantly, how do we want to use the data?
Our merchants are interested in both the current crowd levels and forecasts for the coming hours. They want to access this data online and also be able to download (historical) data. Their focus is primarily on street-level data or for their specific shopping area. There is also considerable interest in visitor profiles, foot traffic patterns, and spending data, although these elements fall outside the scope of VLOED.
There is interest in the following use cases:
Crowd barometer
- For merchants, policymakers, and security services.
- To manage crowd levels effectively.
- Identify areas with lower crowd density.
- Requires detailed spatial data on crowd distribution and forecasts of its evolution.
City guide
- For visitors.
- To guide them through areas of high or low crowd density.
- Needs detailed spatial data on crowd distribution and forecasts of its evolution, with a focus on spatial distribution.
Event overview
- Applying the crowd barometer to events.
- To analyze crowd peaks and their distribution.
- To understand what each event contributes to merchants and the city.
Traffic signage
- Combining the city guide with navigation and route guidance to parking.
- Focus on mobility means this solution falls outside the scope of VLOED.
Open Data sharing
- Focuses on how data is made available.
- Ideally, not only a crowd barometer should be provided, but also the actual data so that merchants (and others) can use it and combine it with their own data.
We also identified a number of use cases related to user profiles, although these are formally outside the scope of VLOED.
Visitor insights
- Types of visitors and their routes.
- For visitor analysis (types) and a visitor guide (tips for visitors).
- Focuses more on data about visitors themselves rather than on crowd levels.
Conversion rate
- The relationship between street crowd levels and cash register receipts.
- To optimize business operations (adjusting opening hours, increasing attractiveness, etc.).
- Involves both crowd levels and visitor types, requiring data from merchants as well.
-
Which measurement tools can we use?
There are various measurement systems and datasets available for purchase. Which ones are useful, and for what purposes?
A living lab in the shopping streets
There are many different measurement systems on the market, ranging from infrared sensors and other types of sensors to cameras with edge computing. Through an innovative public procurement approach, we sought the most suitable measurement systems for our living lab. Various vendors had the opportunity to install their systems in Ghent or Bruges to demonstrate their value and limitations. We are investigating which measurement systems provide the most relevant and accurate data.
In data analysis, we also assess the value of combining these systems with our existing data sources. Additionally, we explore how to make crowd forecasts based on the available data from these measurement systems and our own systems. This is done in a separate living lab focused on crowd forecasting.
The test setups were operational in the spring and early summer of 2022.
Why?
Real-time crowd data and forecasts allow merchants and hospitality operators to make better-informed decisions. With this data and crowd predictions, clear answers to 'crowd-related' questions can now be provided. Through this living lab, data analysis, and the forecasting living lab, we explore what is possible.
For visitors:
- How crowded is the city?
- How crowded are the parking areas?
For merchants:
- When should I schedule extra staff?
- How many people should I expect during the market event?
- Where is the best location to set up a terrace?
- Where is it worthwhile to invest?
How do we measure crowds?
- Radio Sensors in Langemunt and Veldstraat
These radio sensors do not use cameras, Wi-Fi, Bluetooth, or mobile data. They only use a simple radio signal. The more people pass through the signal, the more they disrupt it. Devices across the street detect this disruption. An algorithm counts how many people are in the area based on the signal disturbance. This technology has been used in Ghent since the beginning of the COVID-19 pandemic and is structurally installed in both streets.
- Pedestrian Counting from Langemunt to Kouter
These smart cameras go a step further. They recognize the type of passerby based on their speed: are they a pedestrian or a cyclist? The camera software immediately generates numbers, such as the number of pedestrians, cars, buses, and trams passing through Veldstraat and Korenmarkt. Note: No images are recorded or accessed. Only numerical analysis is transmitted from the system.
What about privacy?
None of the four technologies used in Ghent collect personal data. All tested systems adhere to privacy by design principles. This means we do not seek information about individuals. If any personal data is captured by the measurement systems, it is not allowed to leave the device and must be almost immediately deleted. This ensures that the smart technology cannot be misused by third parties.
-
What data do we already have?
Which existing data can we use, and how does it relate to the measurement systems?
Together with a data scientist, the city has examined the available data. We look at both the output from the living labs for measurement systems and the data we already have.
Various data sources
We have access to a range of data from which we aim to derive valuable insights, including:
- Current and historical crowd measurements.
- Mobility data: street parking.
- Mobility data: parking garages.
- Traffic circulation data from traffic loops and ANPR cameras.
- Other mobility data, such as car-sharing, bike-sharing, bike counters, park-and-rides, Waze, etc.
- Tourism data: hotel occupancy and reservations.
- Visitor numbers for cultural and tourism sites.
- Visitor numbers for events.
The applicability of this data varies significantly:
- (Nearly) continuous data vs. time-limited datasets.
- Data available via API or not.
- Real-time data or post-factum data.
The analysis by VLOED indicates that data is most useful for tracking crowd levels when it involves long time series, available via API, and includes real-time data. Since many data sources are related to crowd levels, this is a key consideration for most systems.
Proxy data for crowds
We identified datasets within Ghent that could serve as strong proxies for city crowd levels. This has been confirmed in further processing for predictive models. It is therefore crucial to focus on the availability and quality of this type of data. Proxy data becomes particularly valuable when combined with data indicating the distribution of crowds within different areas of the city.
Specifically, this includes:
- Occupancy data for parking garages.
- Circulation data based on ANPR cameras.
Data gaps
Every system faces challenges and issues that lead to gaps in the data. This becomes particularly apparent when exploring possibilities for predictive models. For example, a device might fail, or the systems for measuring, processing, or storing data might experience outages. A parking facility might be temporarily inaccessible, or its capacity might be reduced. In all these situations, data and predictive models must accommodate these issues. Therefore, data must be continuously managed and monitored critically.
The main challenge for data analysis and predictive models is handling missing data. Therefore, VLOED also focuses on data monitoring and data imputation.
The need for integrated data management
Multidisciplinary data, long real-time time series, data quality, data imputation, and proxy data all play a significant role in data analysis when actively pursued and managed. To build on the VLOED project, we need a well-developed data management plan. Implementing this plan will also impact the need for specific crowd measurement systems and allow us to work effectively with the data we need.
-
What can we predict?
We can measure how busy a location is right now, but can we also predict future crowd levels based on current data and its evolution?
Thanks to the living lab with measurement systems and data analysis, we now have the input needed to predict crowd levels. We are exploring two types of crowd prediction models, which we will test in both Bruges and Ghent.
Planned tests:
Targeted predictions for a single zone based on measurement systems
- Specific predictions: focused on one street segment.
- Input: data from measurement systems.
- Output: data tested from the same measurement system.
- Uses cases: useful for creating a crowd barometer for that specific area, managing crowds, controlling events, and guiding traffic systems.
Broad predictions for multiple city areas based on multiple datasets
- Wider coverage: distribution of crowd levels over a larger area.
- Input: data from multiple sources, including telecom data indicating crowd distribution, crowd measurement systems, and proxy data.
- Output: useful for a city-wide crowd barometer, overviewing large events, and visitor guidance. Also valuable for analyzing past crowd levels.
Implemented living labs:
Crowd Prediction for a Single Shopping Street
- Forecasts: for one street segment.
- Time frame: for the coming hour.
- Input: measurement system data.
- Output: tested with data from the measurement system, leading to a classification system indicating crowd levels.
Crowd Prediction by Area
- Forecasts: for crowd distribution across different areas.
- Time frame: for the next 4 to 8 hours.
- Input: telecom data, traffic circulation data, parking occupancy, calendar, and weather data. Temporarily trained with measurement system data.
- Output: tested with telecom data. Requires data imputation to address dataset interruptions.
Moving towards production after optimalisation data management
Both living labs show that predictive models are feasible but also have limitations. The model for a single shopping street is achievable, though its added value is more limited. The model with predictions for broader areas is possible but requires ongoing efforts to maintain data quality and retraining. This model highlights the significant value of good proxy data, which requires continuous investment.
Both cities draw the following conclusions:
- There could be a choice to emphasize systems that measure actual crowd levels in various streets.
- There could be a choice for strengthening data management with mobility data that serves as a key proxy for city-wide crowd levels.
-
How can we continue?
What have we learned? And what can we do with it?
Living labs as a method
The living labs allowed us to explore and learn a lot in a short time. We had prior discussions with city services and traders to prepare, and learned that the interests of these actors often differ.
For the technical living lab, we launched an innovative public tender, which allowed us to learn about the measurement systems available.
We learned:
- The potential and limitations of the different systems.
- The essential difference between density and passer-by counts.
- The importance of good locations and set-ups, and what can cloud a good measurement.
- That very many things can interrupt measurements: from power supply and technical breakdowns to, above all, intense use of public space.
- That periodic checks based on manual counts are necessary for each system.
- The peculiarities of each system: each technology interprets slightly differently.
- That this requires close cooperation with suppliers.
The data analysis shows how data can be combined. In particular, the importance of good proxy data is crucial here. To create reliable and useful real-time datasets, it is necessary to build on the existing datasets. The pressure forecasting trial also made us look differently and more consciously at our current data and its usability and limitations.
Data management and predictive models tailored to your city
Other cities
We find that the structure and life in the two cities are very different. Just think of the different impact of tourists and students on Bruges and Ghent, making the dynamics in both cities different. This also makes that the needs for data and systems are different, and pressure forecasts in the two cities also have to be built differently, and the needs for data and systems are also different as a result.
Different data management
Moreover, system and data management is different in the two cities. Ghent has built up data management in which traffic data are crucial, while Bruges attaches great importance to effectively measuring crowds in the streets. Flood has taught Bruges and Ghent a lot for data management; among other things, that the future deployment of data should remain different in both cities.
-
How to work together
Since Covid, everyone discovered crowds. We look over the wall and see what we can learn. And we look for how to share data and information with those interested.
What can we learn from other crowding projects?
Several cities have been working around crowding and pressure data management in recent years. On 27 March, the various cities presented their conclusions at a closing event.
What can we do together?
Data only really pays off when we work together, the data is open, and the necessary background information is available. We therefore want to get to work together with the community. We will communicate more about this soon, but those who do can already get started.
On the Open Data platform, you can already find these datasets. Based on your interest, we can expand this further.
-
