The Norwegian Public Roads Administration (Statens vegvesen) aims to increase tunnel safety by using AI.

The use of artificial intelligence for tunnel event notification provides impressive accuracy, shows one of the pilot projects of Statens vegvesen.

- After the test week in June, we left with very positive - and somewhat surprising - results, says Marius Opheim and Erling Drangsholt in Statens vegvesen, who have recently completed a pilot project in RV. 706 Skansentunnel in Trondheim.

The pilot project "AI for improving AID systems" aims to investigate whether the use of artificial intelligence (AI) can contribute to more accurate warning of serious incidents in Norwegian tunnels.

The system is constantly shouting wolves, wolf 89 Riksvei tunnels use automatic event detection - or AID - as a way to automatically notify the Road Traffic Center of dangerous incidents that occur. About 90 percent of the notifications that VTS receives are fake. Today's AID systems only detect changes in image, not what the change is. Therefore, the system can misinterpret shadows and backlights as unknown moments, and send out false notifications.

-Aid now, the AID system is working towards us, not for us. In order for the system to be more help than a burden for us, we must reduce the number of false alarms, emphasizes VTS operator Drangsholt.

The video equipment in Norwegian tunnels is often from construction year. That may be the reason why today's AID systems are almost accurate enough.

- We wanted to investigate whether the use of AI can help reduce the number of error messages. In addition, we wanted to look at the possibility of reusing as much of existing technical equipment as possible. The results we left with after the test week were overwhelmingly positive, says Opheim.

From 10 to 90 percent accuracy

In a random seven -day period with the old system, barely ten percent of the detections in the "punched vehicle" category were real incidents. After establishing AI, the system had a real detection rate of 90 percent.

- The remaining ten percent were detections of slow -moving cars in rush hour traffic in the morning and afternoon. The system captured cars stood at rest, but misinterpreted as signs of problems. This is something that the AI model needs to be trained to interpret right, explains Drangsholt.

The reason for the significant decline in the number of false alarms is somewhat uncertain. But the pilot's second main finding may have been a major contributor.

- Not a single alarm was triggered by backlight or shadows. This means that we have managed to completely eliminate the biggest cause of error messages. If one manages to reproduce our results in other tunnels, the use of AI for AID can be groundbreaking for tunnel safety, Opheim concludes.

The tunnel's own watchdog AID works by cameras inside the tunnel sending video to a server, which analyzes and defines objects, vehicles and pedestrians. If a possible danger is detected, the system sends a "pop-up" image with direct video current to VTS. The operators then verify whether the incident is real and initiate any measures.

https://kommunikasjon.ntb.no/pressemelding/18009166/kunstig-intelligens-kan-oke-tunnelsikkerheten