Behavioural scanners in Mannheim: testing surveillance that so many cities want (netzpolitik.org)

Mannheim’s “Mannheimer Modell” is a city-scale experiment in automated behavioural surveillance: of 70 public cameras, 46 run pose‑estimation software that converts people into stick‑figure keypoints to classify actions (standing, walking, running, dancing, carrying, fighting gestures like punches, kicks, grabs, defensive postures) and trigger alarms to a police control room. Alarms show a highlighted clip alongside the live camera feed; footage is retained 72 hours or up to 28 days (longer if evidentiary). Developed by Fraunhofer IOSB and in operation since 2018, the system has spread (e.g., Hamburg) and is slated for feature expansion — object/weapon detection and automated face‑search by the state criminal police — while municipal actors push for broader legal grounds to place cameras in so‑called “structural crime” areas. For the AI/ML community the case is a cautionary, real‑world stress test: it raises open technical and governance questions about accuracy, false positives, class and behaviour bias, robustness to occlusion/poor weather, and the limits of pose‑based models for detecting fine motor acts like pickpocketing. It also highlights transparency and evaluation gaps — authorities can’t or won’t publish detection metrics and no independent audit is planned — and legal risks from mission creep (face recognition, lip‑reading) and legislative changes that could normalize high‑intrusion systems without evidence of benefit. Independent evaluation, clear utility metrics and bias testing are urgently needed before such systems scale.