Brighter Cities in Motion

# Ai

# Cities

# Autonomous

# Computer Vision

# Hafnia

# Responsible ai

# Transport

Responsible Computer Vision for Urban Mobility

Cities are where we connect. And when a city’s population doubles, productivity rise by around 15% per capita, according to Geoffrey West and his colleagues at the Santa Fe Institute . At the same time, infrastructure needs per capita fall by a similar amount too. So, cities get brighter as they grow, in multiple different ways.

Michael Porter’s work on clusters points to part of the reason. Proximity speeds up the exchange of ideas, skills and capital. Silicon Valley is the most familiar example. Cities have long worked the same way. Innovation is a contact sport and the density of interactions really matters.

We are in the middle of the AI boom and the climate crisis with both moving fast, and they do not obviously help each other. In this piece we look at one narrow and valuable opportunity inside it. Cities can get better at their core job, which is connecting people safely, smartly and sustainably. They could start by opening their eyes, letting in more light and more visual data.

This article draws on a talk and panel from the recent Transport Innovation Forum 2026 where we explored a simple question: Who has video data they are not making the most of for transport and mobility innovation?

"When transport works, society works." Gustaf Sundell

What cities already have

Cameras are the richest sensors a city owns. Most cities use a fraction of what they could get from them. Two things get in the way. Privacy rules restrict how footage of faces and number plates can be used. And the analytics built on that footage have not always been accurate enough to trust.

Dubuque, Iowa has more than 1,300 cameras. With Milestone Systems and its partner Vaidio , the city now uses them to count vehicles, analyse turning movements, adjust signal timing and spot wrong-way drivers and accidents. It also tracks pedestrians and cyclists. It did this with the cameras it already had. The city reports shorter journeys, faster emergency response and less idling.

Seeing traffic without seeing people

The privacy problem is where brighter AI comes in. brighter AI is a German company that Milestone acquired in 2025. It built a technique called Deep Natural Anonymization. It finds faces and number plates in video and replaces them with synthetic ones. The original pixels are removed, so there is nothing to reverse. Nothing in machine learning is perfect. The GDPR test is whether a person could be re-identified with reasonable time and effort, and brighter AI's method has been independently audited against that.

In one test, a facial search engine quickly found the former German chancellor Angela Merkel in the original footage. In the anonymised version, it found no match.

Blurring and black bars also hide people, but they spoil the footage for analysis. Synthetic replacement, as used in Deep Natural Anonymization, does not. Counting, classifying and spotting incidents all still work. A city can see how its traffic behaves without seeing who is in it.

Deutsche Bahn offers an example. A brighter AI client, it has around 40,000 security cameras. The police keep access to the raw footage for investigations. An anonymised layer could open the same cameras to other uses, such as accessibility and train scheduling. One set of cameras, two sets of rules.

Marian Klee 🔹 , brighter AI's CEO, explains how the technology works in our recent Tech Talk below.

Better data, better answers

Safe footage is only half of the answer. Hafnia is the other half. It is the responsible video data library we are building in partnership with NVIDIA . It holds millions of hours of compliant, curated and annotated video, anonymised with brighter AI. We believe it is the largest of its kind. Developers use it to train and fine tune AI models for traffic and mobility. Better training data is how accuracy improves.

Starting with anonymised data is also faster and cheaper. Once personal data is in the mix, the legal costs and the risk rise with it.

The potential is larger. Our open video management platform, XProtect, runs in more than 500,000 installations. That gives us the potential to access more video data than any other company on the planet. We do not have permission to use all of it yet.

Hafnia also supports vision language models. These let you ask questions of a video and get an answer in plain text almost instantly. At a Copenhagen intersection, we asked one to summarise cycling activity and flag dangerous conditions. At another intersection, just after a collision, we asked it to write an accident report. It filled in the date, time, location and likely cause, in the format an insurer or a police authority needs. The library currently holds about 13,000 accidents. That helps explain how they happen, and how to prevent them.

Genoa shows what this looks like at city scale. In 2025 it became Hafnia's first European engagement. Working with NVIDIA, and with an EU cloud provider hosting the data in Europe, we built a vision language model for transport management trained on Genoa's footage. The work is designed around GDPR and the AI Act. Genoa aims to become Italy's first smart city by 2030.

The data stays with the city. Milestone has a licence to use it in return for a fair exchange of value. Cities using Hafnia get controlled access to the dataset and the models trained on it.

But anonymisation does not replace consent. In a public street, consent is hard to give. What anonymisation does is remove the personal data, such as faces and number plates. Cities still need to say what the cameras are for, who sees the results and how long anything is kept. Without that, the technology does not stick.

Beyond the city boundary

West's finding is about contact. Cities gain because people and ideas collide. The same should hold between cities. The problems are similar from one place to the next, yet most video never leaves the city that collected it.

"We need to stop treating mobility as a collection of isolated systems and start managing it as one connected ecosystem." Willem-Frederik Metzelaar

That matters most for rare events, which no one place sees often enough. Sharing data is one answer. Synthetic data is another. At the Port of Antwerp-Bruges, Hafnia is testing with Dell Technologies whether models trained on synthetic data can help warn of oil spills, using NVIDIA Cosmos. It is early work, but it approaches the same problem from the other side.

Sharing is the route we are working on most. Dubuque, Genoa, Houston and Copenhagen are our pioneer places. They are a cluster of a different kind, linked by what they learn rather than by distance. We have just released our first pedestrian dataset and traffic accident dataset. A challenge is now testing whether an AI model trained in one city works in another. It is harder than it sounds.

The theme ran through the forum. On a panel with Adriana Da Silva Pacheco and Vaidas Paulauskis, we agreed that data only has value when it is shared.

[Insert panel image]

Panel Discussion at Transport Innovation Forum 2026

The catch is that access is not permission, and permission is earned city by city. So the question we ask ourselves daily at Milestone is how to share data responsibly across cities, for safer, smarter and more sustainable places. We are working on it through Hafnia and brighter AI.

If your city has video it is not making the most of, we would love to collaborate. Please get in touch and join our community here .

Brighter cities keep their eyes open. A city should know how its streets work. It does not need to know who you are.

"To develop innovation alone is almost impossible." Paulius Petrauskas

References:

Geoffrey West, Scale: The Universal Laws of Growth, Innovation, Sustainability, and the Pace of Life in Organisms, Cities, Economies, and Companies (2017)