Tunnel & Highway Incident Detection
Real-time vision for road incidents across Turkish highways and tunnels, running on roadside hardware because a datacenter round-trip was too slow.
- İzmir, Turkey — highway incident detection
- Trabzon, Turkey — tunnel incident detection
A vision system watching highway and tunnel camera feeds for the handful of events that matter: stopped vehicles, wrong-way driving, debris on the carriageway, pedestrians where there shouldn’t be any. Alerts go to operators fast enough that they can still do something about it.
It ran on the highway network around İzmir and in the road tunnels of Trabzon. Those are two quite different problems. Open highway means weather and daylight that changes all day; tunnels mean constant sodium lighting, no GPS, and no natural reference for anything.
Accuracy was never the binding constraint. Latency on hardware that was already bolted to the roadside was. We optimised the models through OpenVINO so inference ran on site instead of streaming video to a datacenter and waiting for an answer, and that’s the decision that made real-time alerting possible in the first place.
False positives were the other one. An operator who gets alerted twenty times an hour stops reading the alerts, and after that the system may as well be off. So precision mattered more than recall here, which is not how these models are usually benchmarked.
Dependability
The work ran under FED4SAE, an EU Horizon 2020 programme for getting cyber-physical systems to market, alongside Intel and fortiss GmbH. The dependability side came out of that partnership: how you make the case that a neural network is fit for a safety-relevant deployment, which is a different question from whether it scores well on a test set.
Two papers came out of this work:
- Incident Detection on Junction Using Image Processing
- Application of the Neural Network Dependability Kit in Real-World Environments
ISSD Bilişim Elektronik A.Ş. · 2018–2021