A question from the field
Not every research question is born at a desk. This one was born on the other side of it — in years of volunteer work in civil protection and disaster response, in incident coordination at major emergencies. Anyone who has stood there knows a particular moment: a crowd densifies, the dynamics tip, and the seconds in which you could still do something are short and irretrievable.
Out of that came a question that wouldn't let go of me: can a critical development in a dense crowd be detected early enough to act — without turning the people in it into objects of surveillance?
That is our lab's second research line. In the history of the lab I only touched on it; here I explain why it exists and what stance it carries. We call it CrowdGuard.
The solution we reject
There's an obvious, convenient, and wrong answer to that question: more surveillance. Facial recognition, biometric identification, individual risk profiling — blanket, always-on, across everyone.
We reject that answer on principle, not out of convenience. And in Europe it is, rightly, also legally closed off: the EU AI Act prohibits biometric remote identification in public spaces and, broadly, individual risk prediction. German constitutional law draws a clear line through the freedom of assembly — those who are watched no longer behave freely.
A system meant to make public spaces safer by capturing every individual pays for that safety with the very freedom that makes a public space worth having. That is not a trade we make. For us, civil liberties are not an obstacle to build around — they are the design principle you build with.
The physics of the crowd, not the face of the individual
If the identity of the people is off-limits — what's left? Surprisingly much. Because danger in a dense crowd is rarely a question of who is standing there. It's a question of physics.
A crowd has measurable properties that have nothing to do with identity: its density (people per square metre), its direction of flow, points where movements stall against each other. Safety research has known the critical thresholds for decades — the work of Fruin and Still describes the density at which a crowd can no longer control its own motion and becomes dangerous.
That is exactly where we start: not at the individual's face, but at the dynamics of the whole. The system estimates density and flow from the image — and outputs anonymous, deliberately coarse skeletons, not identities. It doesn't ask "who is this?", it asks "is this area about to become critical?". The difference isn't cosmetic. It decides whether a tool serves protection or becomes surveillance infrastructure.
Forensically grounded — and with restraint
A method like this must not rest on guesswork. Our validation basis is publicly documented incidents from the 2015–2025 period, selected by the density of their forensic documentation — that is, by how well-traceable what actually happened is.
And here belongs a sentence that matters to me: in this work we use no specific case names. Out of respect for the bereaved, for ongoing proceedings, and for places of remembrance. Behind each of these incidents are real people and real suffering; they are not marketing material and not a demonstration object. The selection logic is presented methodically and soberly in a separate companion text, not spread out prominently in an overview. Research on a subject this serious demands restraint — and restraint is not weakness, it is part of the method.
Two tools that run offline
CrowdGuard today is a research workspace, not a product — but a concrete one. Two standalone, fully offline-capable tools show where the line is aimed:
- An Inspector that demonstrates movement and density analysis from real footage — with honest, anonymous skeletons instead of identities, and with openly stated limits where the estimate is uncertain.
- An incident-coordination situation map for command-post work, modelled on the logic of the German FwDV 100 — the shared situational picture on which an incident command makes decisions, georeferenced and cloud-free.
Both run locally, with no data outflow. That is the same stance on data sovereignty that carries our whole lab — here it isn't optional, it's a condition.
Honest research means reporting failures too
I distrust safety claims that only know successes. So this line carries an uncomfortable discipline: we also report what does not work. An approach that sounded promising was honestly discarded at a hit quality near chance, rather than being massaged into looking good. More than six hundred automated tests keep the state of the tools verifiable.
That is slower than a catchy promise. But for a system meant to support emergency responders when it counts, a flattered number is more dangerous than an honest gap. Those who carry responsibility must know what they can rely on — and what they can't.
Why a memory lab protects public spaces
At first glance it's a leap: from the long-term memory of artificial systems to safety in a crowded square. At second glance it's the same conviction, just in a different field.
It's the conviction that a technology must be traceable, data-sovereign, and aligned with people — and that you don't weigh civil liberties against utility, you build them into the architecture. In ZenBrain that means a memory you're allowed to look inside. In CrowdGuard it means protection that works without surveillance. Different fields, the same line.
CrowdGuard is an independent, self-funded research effort. I keep explorations with innovation networks and public-sector partners strictly in the conditional until something binding is in place — I name names only after written consent. More on the current state is on the public-spaces page.
What remains
The convenient narrative says safety and freedom are opposites, and you have to choose. I think that's wrong — and I think it's a dangerous convenience. The hard, worthwhile work isn't trading one for the other; it's building both into the same system.
That is exactly the ambition of this research line: a public space that is safer because it does not surveil the people in it. That's harder than a camera system with facial recognition. But it's the only kind of safety that lets a public space stay public.