How we built crowd counting models for unticketed events – and how government is using them
Our Director of Public Services, Paul Maltby, explores how our crowd counting models for unticketed sporting and cultural events are currently being used by government to evaluate impact, return on investment, and future funding decisions.
The rapid evolution of generative models has thrust AI into the mainstream debate about the future of UK public services. At Faculty, we’re increasingly focused on the power of and agentic systems that can ‘act’ to solve administrative problems, and their potential to improve how government works. But amid all this, it's easy to forget that more traditional machine learning deployments are already delivering useful tools for government.
Predicting event attendance with the DCMS
An example of these successes is a recent project we delivered for the Department for Culture, Media and Sport (DCMS), in partnership with social researchers at Verian and academic collaborators at the University of Glasgow, to develop new models for estimating attendance at unticketed cultural and sporting events. This is a long-standing challenge in government: the absence of reliable data creates clear difficulties for policy questions such as evaluating impact, return on investment, and future funding decisions. Traditional methods often struggle to capture hyper-local participation, particularly for short-duration or open-access events, limiting their ability to fully understand engagement across the cultural and sporting sectors. For our data scientists, it was a brilliant opportunity to sink their teeth into a classic ML problem to help bridge that gap.
There are all sorts of untapped privacy-compatible datasets out there that government could be leveraging for ML. We identified more than 50 across transport, geo-location apps and social media. The likes of Strava even have dedicated self-serve platforms available to departments, providing anonymised and aggregated crowd mobility data from running and cycling events. We used these to develop novel modelling approaches, each utilising a different core data source. These approaches were tested against a variety of events, including the Bradford 2025 UK City of Culture series and the Great North Run, as well as ‘always open’ sites like the Giant's Causeway and the British Museum.
Beyond testing the operational feasibility of these new crowd counting methods, the project developed methods to assess viability against a range of ethical and safety criteria, going beyond traditional PII issues. The thinking and approach to risk in this space is maturing rapidly.
A closer look at one method
One of the approaches we took was running computer vision models over drone footage to count crowd attendance. It’s been fascinating to see how the proliferation of open source packages – models like YOLO-CROWD and DMcount for those interested – continue to accelerate the development cycle for these types of solutions. We were able to rapidly return good results where high-quality drone footage was available, such as in an event series during the Bradford 2025 UK City of Culture, where crowds gathered for public processions.
Inevitably, this approach is more complex than just running an ‘out of the box’ model. We had to carefully extract individual scenes that were representative of the event and aggregate counts for each frame to arrive at an overall prediction. This involved applying another technique called SIFT (Scale-Invariant Feature Transform) to every frame to identify consistent features (i.e. human attendees), even when the perspective changed across shots. Using these features, the models created polygons around previously identified individuals, preventing double-counting as the analysis progressed through the frames. This approach ensured an accurate total count for each scene by tracking attendees (importantly without ever identifying their person) across all frames, rather than relying on isolated snapshots.
These types of novel approaches have allowed DCMS to model attendance across a range of events. Amidst the latest excitement with agentic AI, it's worth remembering that it’s not all about agents. The proliferation of open-source datasets and more traditional AI packages is lowering the barrier to applying AI to routine government tasks, and our crowd-counting models are currently in the hands of DCMS data scientists to explore this.
A full write-up of this work is published on Gov.UK: https://www.gov.uk/government/collections/rd-science-and-analysis-programme-capturing-engagement-numbers
