Germany · de.indymedia · 3h
Cameras and drones in Berlin: What behavioral scanners are supposed to detect
Deutsch (original) · Auto-translated to English
Fixed and mobile police camera systems are intended to monitor endangered objects in Berlin and permanently scan everything that moves around them. Drone recordings could also be combined with behavioral scanners. This emerges from a service description that has now become public.
Thehttps://www.tagesschau.de/inland/faq-hackerattack-berlin-follow-100.htm...from the beginning of August to the Berlin administrationand the subsequent release of almost one and a half million files also provides insights into police surveillance projects using object recognition and behavioral scanners. Because among the documents published by the attackers is thishttps://cdn.netzpolitik.org/wp-upload/2026/09/Berlin-KI-Objectschutz-Lei...to the “Project to test AI-supported video technology”(pdf). This formulates in detail the technical requirements that the contractor must meet for the desired Berlin police computer system.
To put it simply, the surveillance system should search through recordings from cameras in predefined areas of public space and according to predefined criteria and sound an alarm if it detects conspicuous objects or movement patterns. The service description lists which suspicious behavior and movements should lead to an alarm.
In all suspicious situations, a police officer should look at the automatic pattern recognition recordings and decide what happens next. Such a follow-up measure could, for example, be storing the recording in a legally secure manner or sending police colleagues to arrest someone who has been recorded.
It's not about the behavioral scannershttps://netzpolitik.org/2026/kritikitaetstreib-in-berlin-problemat...so-called “crime-prone places”, but this time about police property protection, specifically the Berlin Red Town Hall, the House of Representatives and the headquarters of the Senate Department for the Interior and Sport. Therehttps://taz.de/KI-Videoueberwachung-in-Berlin/!6160851/">willthe police “AI-powered” surveillance systemstest before they go into regular operation.
The desired details of the automated evaluation of the video recordings are listed in the service description for these buildings. The fact that the House of Representatives is also on the monitoring list was not without criticism:https://taz.de/KI-Videoueberwachung-in-Berlin/!6160851/">NiklasSchrader from the Left Partycalled it an “affront to Parliament” if MPs became “guinea pigs for the police’s AI experiments”.
Mobile camera systems and drones with behavioral scanners
Self-learning software is not provided for in the service description, but rather the “human-in-the-loop principle”. This means that the systems should not be able to independently draw any conclusions from recordings that have already been made. The restriction is probably due to the high demands that the Federal Constitutional Court has placed on the protection of fundamental rights in self-learning systems. The high degree of interference with the basic rights of all passers-by is probably the reason why biometric facial recognition should not be used either.
In autumn 2025, the police sought to speak to those responsible for the buildings, as monument protection issues must be taken into account when attaching video cameras to the building shell. At the same time, the service description also includes mobile camera systems for monitoring the Berlin House of Representatives. Even drones are already planned, but are described as “optional” in the document.
The service description shows that the recordings from the mobile cameras should be able to be integrated into the overall system and work “with identical alerting rules”. The video stream from drones should also work with the same functions as stationary solutions.
From a technical point of view, recognizing the behavior patterns for alarms on mobile cameras and drone recordings from above or from diagonally above is likely to place new demands on the software. Because moving camera systems or the top view naturally change the patterns that need to be evaluated. In addition, “temporary or ad hoc cameras” should also be integrated.
The planned start of construction on the buildings was September 1, 2026. The system is scheduled to start up at the end of 2026, in “cooperation with the Fraunhofer Institute”. This probably refers to the Fraunhofer IOSB (Optronics, System Technology and Image Evaluation) from Karlsruhe, which has been working with the company for almost a decadehttps://netzpolitik.org/2026/behaviourscognition-in-mannheim-10-000-film...Police testing behavioral scanners. How successful the long-term project there is and what error rates the police officers have to deal with is not publicly known.
https://netzpolitik.org/2026/behaviourscognition-in-mannheim-10-000-film...Films for a Hallelujah
What should the system detect?
It has not been conclusively defined which moving objects and with which parameters the system in Berlin should recognize and evaluate. The service description only lists examples such as people, types of vehicles, colors, headgear or bags and backpacks.
However, it lists the scenarios that should be recognized. Objects left behind that are larger than 30 centimeters and have been lying around for more than 30 seconds must be identified. The system must also be able to detect illegally placed objects in previously defined security areas. It should also be mandatory to detect damage to fences or gates.
Other detection services specified in the service description are not defined as must-have features, but are desired. Entering restricted areas, crossing virtual lines, climbing fences and walls and “tailgating”, i.e. when one person or object moves closely behind another, should trigger an alarm.
In addition, it would be desirable to also detect loud noises that are above 85 dB(A) for more than ten seconds. The system should also, if possible, alert if certain acoustic patterns indicate the use of firearms or detonations.
As far as the behavior of people in the area of surveillance is concerned, throwing objects and spraying/smearing actions should be detected. Light projections are also intended to trigger an alarm.
The behavioral problems that can be detected in people are finally listed, but are not marked as must-have features of the system:
- Staying/walking around for a long time without reason,
- Long-lasting photography,
- attaching/removing objects,
- Acts of theft […],
- Repeated appearances (without biometric identification).
False alarms inevitable
Since such behavior and pattern recognition is always subject to errors, the false alarms recorded from an initial three-month test phase should be used to improve the system as much as possible. The service description explicitly provides for this. Only then should the nine-month “real trial operation” begin.
However, relearning is not an easy undertaking, as real recordings of people would be used. Their basic rights must be respected; they cannot be used as input material for AI training without being asked.
In addition, the prohibition of discrimination must be observed. For example, people with physical disorders, autism or physical disabilities often move differently and therefore noticeably in the sense of technical pattern recognition. However, they must not be discriminated against.
The police solve the problem by simply specifying “ensuring non-discriminatory algorithms”. There is a footnote attached: “It must be explained how this will be achieved.”
When it comes to error rates, however, things are becoming more precise: “AI-supported video analysis must operate with an overall recognition accuracy of at least ≥ 95%.” This refers to all types of object and pattern recognition in the software. The service description further stipulates: “The maximum average false alarm rate should not be higher than 25% across all scenarios.”
These specifications are surprisingly low for pattern recognition systems. In practical terms, this means that the software can make many errors and therefore constantly issue numerous false alarms, but also leave objects or behaviors that need to be detected undetected. For individual scenarios, over a quarter of all alarms may be false; on average, the false alarm rate may not exceed 25%.
Whether the stated goal of relieving the burden on the police when it comes to protecting property can be achieved with the software naturally depends on these requirements. The false alarm rate specified at the beginning should, if possible, decrease after the three-month test and “retraining”.
What should not decrease, however, is the number of buildings where the cameras with behavioral scanners can be used. Because the Berlin Interior State Secretary Christian Hochgrebe (SPD) has alreadyhttps://taz.de/KI-Videoueberwachung-in-Berlin/!6160851/">angekterminates, also taking a look at embassies and other endangered places.
How well the system really works after a year of testing and “real trial operation” is likely to be of interest to the public because of the planned expansion of monitoring. But there is a high probability that the results will be hidden. It is already known from the long-term test carried out by the police in Mannheim that once systems have been put into operation, they continue to run for years without the results being published or even independently evaluated.
But perhaps the upcoming election in Berlin could set the course differently: the Greens, the Left and the FDP want to limit video surveillance. Like the BSW, the SPD is also rather moderate on the issue in its election program. However, as the current government party, the SPD has pushed for the expansion of legal powers in the amendment to the General Security and Order Act (https://gesetze.berlin.de/bsbe/document/jlr-NNLBE00004772NN00000000055">...) and thus makes the behavior scanners in the police law possible. If, on the other hand, the CDU and the AfD were to win a lot of votes, it would be their declared intention to expand video surveillance quantitatively and qualitatively.
The outcome of the election and the subsequent formation of a government is still uncertain. After the IT attack on the Berlin administration, it is also uncertain whether IT security in the capital can be quickly changed from inadequate to presentable. This could contribute to the results of the behavioral scanner test becoming unintentionally public at some point.
The work of netzpolitik.org is financed almost 100% from donations from our readers.
Become part of this unique community and support our public interest-oriented, advertising- and tracking-free journalismhttps://netzpolitik.org/spenden/?via=rss">nowwith a donation.
Read the full story at the source →
Source: de.indymedia