Smart-Tunnel & Highway Incident Detection
Deep-learning incident detection for tunnels and highways, including the FED4SAE Smart-Tunnel project with Intel and fortiss.
- 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.
Smart-Tunnel · FED4SAE
Smart-Tunnel is the FED4SAE-supported deep-learning enhancement of ISSD’s automatic incident detection system. It detects pedestrians and stopped vehicles in tunnel camera feeds, using Intel Movidius Myriad X VPUs in a high-density deep-learning (HDDL) configuration.
FED4SAE reports that a live customer evaluation in Turkey roughly halved false alarms compared with the legacy engine, while detecting every event that the earlier engine detected and raising alerts faster. Hardware cost per camera stream also fell. These results describe that evaluation, rather than a guarantee for every deployment.
Intel supported hardware benchmarking and inference tuning; fortiss supported robustness assessment with the Neural Network Dependability Kit. Blumorpho provided business coaching. The engine was designed for both new installations and retrofits to existing tunnel monitoring systems.
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