What this is about
At the junction of Grafenberger Allee and Simrockstraße in Düsseldorf, cars repeatedly turn left where it is forbidden — across the tram tracks, right in front of approaching trams. For passengers on board, for the drivers and for everyone else at the junction, that is a real danger: emergency braking, damage to property, and in the worst case injuries.
Until now, nobody knew how often it actually happens. There were observations from tram drivers, there was gut feeling, there were individual documented incidents — but no reliable number. Without a number you can neither justify changing something nor verify whether a change worked.
That is the gap this project closes. Together with us, Rheinbahn is trialling a system that counts the violations. It has been running at this junction since August 2026.
What is installed there
Two traffic cameras and a small computer in an enclosure on a mast. The cameras look at the junction area only — the way a person at the roadside would, making tally marks on a sheet of paper.
The computer detects vehicles in the live image, follows their path for a few seconds and decides whether a turn was illegal. Trams and buses that are allowed to turn there are recognised as exempt and not counted.
What the system records is a tally entry: timestamp, vehicle category ("car", "truck"), direction, classification. Nothing else.
What happens to the images
No video is stored and no images are transmitted. This is not a promise of good behaviour, it is built into the architecture: all processing happens on the device at the junction. The camera images exist for fractions of a second in memory, are evaluated and immediately discarded. During normal operation the device has no way to persist image material or send it anywhere.
In concrete terms:
- No licence plate recognition and no face recognition. The software distinguishes vehicle categories — car, truck, bus, motorcycle, tram. Identifying a specific vehicle or a specific person is not part of the design and cannot be retrofitted without rebuilding the system.
- No fines, no reports. The system counts anonymously. It reports nobody to the authorities, and the data it collects cannot be traced back to a vehicle owner.
- No re-identification. Vehicles are tracked through a single turn under a random number that expires after a few seconds. Nothing is linked across manoeuvres.
- Processing in Europe. The anonymous counts are stored encrypted with a European provider inside the EU. No US provider is involved.
Two exceptions exist for setup and quality assurance, both triggered manually by a human: single still frames to configure the detection zones, and short video clips to verify counting accuracy. They stay local, are deleted after evaluation and are not part of regular operation.
The goal
What should come out of this is a statement you can act on: how often illegal turns happen, when — rush hour, darkness, weekends — and how that changes once something at the junction changes.
That serves two purposes. First, it substantiates the hazard instead of asserting it: a measured number is a different argument than an impression. Second, it makes the effect of countermeasures verifiable — clearer signage, different signal timing, structural changes. Whether and which measure follows is for Rheinbahn and the responsible authorities to decide; the system only supplies the numbers.
Operation is time-limited. Whether it continues will be decided on the basis of the anonymous counts.
How it works technically
- Everything on the device. A Raspberry Pi compute module with a Hailo-8 AI accelerator sits on the mast and evaluates the camera stream in real time. Only counts leave the site.
- Off-the-shelf detection, our own domain logic. For detecting vehicles we use an established YOLO model. The actual work sits in the layer above it: zones and geometry of the junction, heading, and classifying a manoeuvre as violation, suspected violation, U-turn or exempt — including how to handle occlusion, darkness and rain.
- Resilient without connectivity. If the mobile link drops, the device buffers counts locally and uploads them later. Nothing is lost, nothing is counted twice.
- Analysis in the EU. The numbers land in a PostgreSQL database on European infrastructure; a dashboard shows trends by time of day and weekday.
Who is behind it
The project is operated by Rheinbahn AG as the controller for data processing. Technical implementation is by schnaq GmbH in Düsseldorf. Questions about the measure itself are answered by Rheinbahn; for technical questions about the system, talk to us.
Why this work suits us
Traffic research with cameras is a data protection minefield. Anyone working here without a clear architecture fails not on the technology but on public acceptance. We bring both: computer vision on edge hardware — and a sober view of which data should be created in the first place.