Track B

Safety in public spaces — without biometric identification

„How can AI detect dangerous situations early without violating the fundamental rights of uninvolved people?"

A safety architecture that detects collective anomalies rather than classifying persons. Civil liberties by design, not as an add-on.

Motivation

Why this research

There is no off-the-shelf technical solution in the space between security and fundamental rights. Classical predictive-policing systems operate with person-level risk scores; Art. 5(1)(d) of the EU AI Act prohibits exactly that. Art. 5(1)(h) likewise restricts biometric remote identification in public spaces, with narrow exceptions.

The question, therefore, is not how to circumvent these prohibitions, but what an architecture has to look like to be useful under them. This research answers with a methodology that models collective dynamics instead of individuals — and that makes protecting helpers (rather than classifying suspects) the explicit objective.

The applied orientation of this track is informed by first-hand operational experience in civil protection and disaster response — where the realities of the field act as a check on purely theoretical assumptions.

Methodological approach

Three-outlier model and helper protection

Instead of deterministic person classification, the three-outlier model produces exclusively calibrated probabilistic indications of statistical anomalies in collective dynamics. No person-level labels, no biometric identification, no suspicion generation — only an indication system for human decision-makers.

The second central pillar is helper protection as the objective: the system is optimised to detect and protect helpful behaviour rather than to produce suspect cases. This reversal of the objective changes the whole architecture — from annotation to threshold calibration.

Processing principle

From a collective pattern to a calibrated indication

The processing chain is designed to preserve fundamental rights: aggregated movement and density patterns yield a calibrated, probabilistic indication — not a person label. A human always makes the decision.

Collective dynamicsaggregated movement patternsAnomalystatistical, in the patternCalibrated indicationno person labelHuman decideshuman-in-the-loopAggregated patterns, not identities — no biometric identification, no person tracking.
Collective dynamicsaggregated movement patterns
Anomalystatistical, in the pattern
Calibrated indicationno person label
Human decideshuman-in-the-loop

Aggregated patterns, not identities — no biometric identification, no person tracking.

The model, visualised

Three roles, one protected

The three-outlier model distinguishes roles in a crisis situation rather than classifying individuals. Helping people are structurally protected; fundamental-rights guardrails are written into the architecture.

SourceinitiatingprotectedHelperintervening, assistingKnowercontextually reactingCIVIL-LIBERTIES GUARDRAILSNo biometric identificationNo real-time person trackingNo facial recognition
Sourceinitiating
Helper · protectedintervening, assisting
Knowercontextually reacting
Civil-liberties guardrailsNo biometric identification · No real-time person tracking · No facial recognition

Core components

The civil-liberties architecture at a glance

Three-outlier model

Distinguishes three roles in a crisis situation: source (initiating), helper (intervening, assisting) and knower (contextually reacting bystander). The role differentiation structurally protects helping individuals from being misclassified as initiators.

Helper protection as objective

Optimisation not for suspicion but for the detection and protection of helpful behaviour. Inverts the detection logic.

No biometric features

Annotation schemes are designed to avoid facial, gait, and voice features. Data minimisation as an architectural decision.

Eight mandatory corrections

Eight methodological corrections defined in the research-ethics document that every implementation of this architecture must satisfy. They constitute the civil-liberties architecture.

Probabilistic calibration

Probability statements are calibrated against publicly available recordings of documented incidents. No training on personal data without a research clause.

External validation

Notified-Body pre-assessments and legal opinions on the fundamental-rights conformity of the architecture are foreseen. The bodies under consideration are named on the Research Ethics page.

Legal frame

What the architecture rests on

  • Reg. (EU) 2016/679 (GDPR) Art. 89 in conjunction with § 27 BDSG — research clause
  • Reg. (EU) 2024/1689 (AI Act) Art. 5(1)(h) — ban on biometric remote identification
  • Reg. (EU) 2024/1689 (AI Act) Art. 5(1)(d) — ban on personal risk prediction
  • Reg. (EU) 2024/1689 (AI Act) Annex III — classification as high-risk system
  • BVerfGE 69, 315 — Brokdorf line protecting political assembly

Get involved

Collaboration on this research

Universities with a public-safety focus, public authorities at the BMI/BSI/BBK level, and research consortia along BMBF SIFO and Horizon Europe Cluster 3 are explicitly invited to get in touch.

public-sector@zensation.ai