Germany de.indymedia
Behavior Analysis Training Videos: There is never enough violence for surveillance AI
The Berlin behavioral scanner is trained with AI videos. This is a challenge for the quality of the software. In order to improve it, additional data from monitored people is used.
Deutsch (original) ·Auto-translated to English
The Berlin behavioral scanner is trained with AI videos. This is a challenge for the quality of the software. In order to improve it, additional data from monitored people is used.
The Kottbusser Tor on a weekday afternoon: people inspect the displays at the vegetable stand. A woman carefully comes out of the Rewe, carrying two shopping bags in each hand. A small group is having a heated discussion while one person waves a piece of written paper. Police officers in stab-proof vests patrol the square.
All of these people are currently being analyzed by software. 30 cameras film the eventhttps://goerli247.org/2026/11/01/ki-kameras-am-kotti/">onthe northern half of the squareand so-called AI checks who is doing what and whether what is happening is evidence of a crime. The automated continuous monitoring is intended to enable control at a previously unknown level of efficiency.
In order for the software to recognize which type of scenes it should raise an alarm for, many clips were previously fed in. In this training material it must be marked whether the scene is intended to be an example of behavior to be tolerated or of an alarm to be triggered.
Not enough violence
The amount and quality of training data has a big impact on how well the system ultimately works. The central problem at this point: Behavior recognition can learn very well what it looks like when a person walks using images from public spaces - there is a lot of visual material there. But the factors that can be used to identify a fight are difficult to categorize based on what is happening on the ground. From the perspective of AI trainers, there is simply not enough violence in public spaces. “As became apparent during the course of the project, unfortunately there is only very limited public data available,”https://netzpolitik.org/2025/behavioursscanner-im-mannheim-hier-wird-die...the Mannheim police.
That's why the AI developers work with images of inauthentic physical confrontations. In Mannheim, for example, police officers have already organized 2,000 show fights for AI training; the Mannheim police assume thathttps://netzpolitik.org/2026/behaviourscognition-in-mannheim-10-000-film...another 8,000 fights neededuntil the software is suitable for practical use.
The Mannheim training videos are classified as “classified information” and are also subject to personal rights regulations, which is why the Mannheim police do not want to release them. However, one clip is public; it is shown in Mannheimhttps://www.technoseum.de">Technoseumshown as an example for the training videos. We filmed it there.
–––
Against ID checks online.
We are fighting for an open internet.https://netzpolitik.org/spenden/?via=rss">Thisonly possible with your support.
–––
https://cdn.netzpolitik.org/wp-upload/2026/11/XMannheim.mp4?_=1">href="https://netzpolitik.org/wp-upload/2026/11/XMannheim.mp4">https://cdn.netzpolitik.org/wp-upload/2026/11/XMannheim.mp4">https://net...
The police officers that can be seen appear sedate, and the suggested punches and kicks are reserved. In the image you can also see the stick-figure abstraction of the movements that the AI uses to categorize them.
AI teaches AI how people fight
The developers of the software that scans passers-by at Kottbusser Tor in Berlin are taking a different approach. In Berlin, no police officers should be required to take part in show fights. Instead - in addition to some real scenes - AI-generated videos were largely used for AI training. The surveillance software should learn what violence looks like from a video generator that imagines presumably violent scenes. “In this way, the amount of training can be increased systematically and in compliance with data protection regulations, without having to rely on location-specific data,” writes a PR person on behalf of Staige, the manufacturer of the monitoring software.
We have made an IFG request for the complete training data set and hope to be able to use it for a movie night soon. Until then we have to make do with a trailer: On thehttps://staige.com/">Websiteby StaigeUnder “Violence Detection” you can see three AI-generated clips in which AI appears to have detected violence. In fact, the figures depicted do not seem to touch each other at all. Other oddities catch the eye: tram doors open at full speed. People smile and wave at surveillance cameras or hop like bunnies. There are no colorful stick figures to be seen; the mathematical abstraction with which the software analyzes what is happening remains hidden.
Heiko Paulheim is a professor of data science in Mannheim, where a behavioral scanner has been recording public spaces using countless cameras since 2008. How good such systems are, says Paulheim, depends heavily on the training data. If they are not authentic, it affects the quality of the software.
“Like children’s drawings of cats”
"You can think of it a bit like image recognition for cat pictures that is trained on children's drawings but applied to photos. The performance will always lag behind a model that was trained on photos," says Paulheim. To do this, the video generator can write distortions into the training data: “Models that are trained on YouTube videos produce material that is similar to YouTube, possibly with a tendency towards a lot of action, good lighting, a focus on younger actors with popular body shapes – and perhaps also increasingly common stereotypes such as that of the dark-skinned attacker in the black hoodie.”
According to Paulheim, this can lead to two types of errors. First: problems are not recognized. “If I only have action-packed fights between young men, I might not recognize the comparatively low-action snatching of an older woman’s handbag,” says Paulheim.
–––
The work of netzpolitik.org is financed by donations from our readers.
You too can support our public interest-oriented, advertising- and tracking-free journalism.https://netzpolitik.org/spenden/?via=rss">Donatenow..
–––
The other type of error is the false positive. “The software recognizes a scene with a dark-skinned person in a hoodie as problematic, even if this person is not behaving in a problematic way, but is just jogging through the image, for example.”
Berliners provide free training data
A PR person who works for Staige gives us more details about training the surveillance software. “The training material includes scenes of exchanges of punches, kicks, violent pushing or pulling movements with ground contact, and scuffles at ground level,” he writes, for example. To achieve this, some of the movements of actual people were recorded using motion capture technology. The video generator then created a large number of clips from the abstracted event, which show the process with different people at different times of the day. According to the Staige spokesman, the AI clips are created “on site”, i.e. in a Berlin police property.
In addition, the Berlin model will be “further optimized” with the movement data of the people who are at Kottbusser Tor, according to the PR person. The police officers who observe the videos then note whether an alarm from the surveillance AI was justified or not. And the software then also uses this scene as an example for “ignore” or “alarm”.
The software is fed with the movements of people at Kottbusser Tor. The movements - for example the gait - are a unique identifier, as unique as a fingerprint. But with the behavioral analysis, the outgoing black-red Berlin state governmenthttps://netzpolitik.org/2025/neues-polizeigesetz-berlin-wurf-die-freihe...last Decemberalso allows surveillance AI to be trained with personal data – and this data to also be passed on to the software manufacturers.
–––
Your donation for digital freedom rights
We report on current internet policy developments, uncover scandals and initiate debates. We are completely independent. Because our fight for digital freedom rights is financed by donations from our readers.https://netzpolitik.org/spenden/?via=rss">Nowdonate.