Algorithms on Patrol: Bremen's Trams Become an AI Surveillance Zone

Contents

In Bremen's trams, the gaze will in future no longer wander only out the window, but also from the ceiling down with digital precision onto what is happening in the aisle. Between nine and eleven camera eyes are already watching over passengers per car. But by the end of 2026, there should be far more behind these lenses than just passive recording: Bremer Straßenbahn AG (BSAG) plans the widespread use of Artificial Intelligence (AI). The technology is intended to recognize potential dangerous situations, aggressive behavior, and acts of violence the moment they arise.

The program now announced, according to the plan, which taz reports on, uses the existing camera infrastructure of the vehicles and upgrades it digitally. The system is called AI-Watch and comes from the Bremen-based software company Just Add AI.

Initially, more than 40 of BSAG's trams are to be equipped with the technology in the course of the year. Bremen aims to take a pioneering role nationwide in public transport. While other municipalities are still debating ethical hurdles, the Hanseatic city is taking action. AI-supported camera systems in public transport have not existed in any other federal state so far, according to BSAG.

Alarm: When AI Calls the Police

The heart of the innovation is real-time analysis: if someone is being insulted, harassed, or attacked, AI-Watch is supposed to sound an alarm at BSAG's control center. If the algorithm detects a pattern that indicates escalation, it immediately sends a pulse. This triggers a signal in the driver's cabin and switches the live images directly to the control center. The final decision-making power still lies with humans there: employees assess on the monitor whether there is actually a threat and whether the police need to be alerted.

"This allows for faster action," emphasizes Bremen's Senator for Transport, Özlem Ünsal (SPD). A crucial advantage is also that drivers can react to incidents that would have remained hidden to them without the digital support behind them. The project is costing the stakeholders a lot: around 10,000 euros per tram, with the Federal Ministry of Transport covering 80 percent of the total costs. Given this funding, it is hardly surprising that, according to a BSAG spokesperson, other public transport companies have already expressed interest in the program.

Subjective Sense of Security vs. Objective Danger

The necessity for such a system is often justified by a vague feeling of insecurity. Thorsten Harder, BSAG's Chief Technology Officer, knows: "I have noticed a changed subjective perception regarding the sense of security in our vehicles." Although the trams are already safe places. However, serious incidents like a queerphobic assault in December, in which two young women were beaten unconscious shortly before the main train station, show the limits of the current surveillance. In precisely such situations, AI is intended to enable faster intervention in the future.

From a data protection perspective, the project is allegedly on solid ground: a certificate of non-objection from the state data protection authorities is said to be available. To protect privacy, data processing takes place directly in the vehicle. Personal data is not used to train the AI's accuracy. As soon as an image is transmitted to the control center, the system automatically pixelates the faces of those recorded. However, characteristics such as clothing, skin color, or gender remain recognizable to enable identification in an emergency. In parallel, a local data carrier stores the image material for 72 hours in non-anonymized form so that the police can access it specifically in case of criminal offenses.

Error Proneness: Hug or Punch?

The technical challenge lies primarily in the error rate. The fact that camera systems still have a lot to learn is shown by a look at Hamburg, where AI surveillance is already in use at Hansaplatz. There, the system has already mistaken fights for warm hugs in the past. The Bremen officials, however, claim that AI-Watch works more precisely. The solution was trained in "theater performances" where actors simulated attacks. A special feature of the algorithm is the analysis of the environment: the AI not only looks for potential perpetrators but also for the reactions of bystanders.

If passengers turn around in fright, for example, the chance of the AI sounding an alarm is higher. Nevertheless, the technology remains prone to errors. In an initial pilot phase, which began in April last year, the AI was initially tested without real-time transmission. It identified 16 potential danger situations. However, a comparison with official operational reports confirmed only four of them. The remaining twelve reports turned out to be false alarms or missed alerts. Until all planned vehicles are equipped, there remains room for the question of whether the digital upgrade actually offers the promised protection or primarily increases the density of surveillance.

(nie)