Incident detection in Tunnels in Norway

Challenge:

The existing automatic incident detection (AID) system in Norwegian tunnels has been underperforming due to outdated video technology and basic image recognition capabilities. These systems detect changes in the visual environment, but they lack the sophistication to recognise actual hazards. As a result, approximately 90% of the alerts generated by this system are false alarms, overwhelming traffic operators with unnecessary notifications. This presents a significant challenge, as the high rate of false alarms not only burdens the operators but also diminish the system’s credibility. 

Approach:

To address this issue, Statens Vegvesen decided to run a PoC using computer vision, sensor data and machine learning to enhance the AID system’s precision and reliability. The end objective is to develop a system that can be trained to distinguish between real incidents and irrelevant changes in the environment with greater accuracy. Deep learning models can analyse vast amounts of data from various sensors and video feeds, learning to identify patterns and anomalies that signify genuine incidents. Additionally, these models have the capability to learn continuously, improving the system’s performance over time as it adapts to new data and evolving conditions. Statens Vegvesen initiated a pilot project in the Skansentunnel in Trondheim, integrating AI with deep learning capabilities to enhance the existing AID system. This project focused on using existing tunnel video and sensor equipment to minimise costs while improving detection accuracy. AID works by cameras inside the tunnel sending video to a server, where the computer vision techniques analyse images to detect objects, vehicles and pedestrians and machine learning is used to assess the danger. If the situation is qualified as dangerous, the system sends a "pop-up" image with a direct video stream to the system operator. The operators then verify whether the incident is real and initiate any measures deemed necessary.  

Outcomes:

The AI system was evaluated for its ability to accurately detect and categorise incidents over a test period, significantly reducing false positives and increasing system efficiency. The integration of AI into the AID system led to an increase in detection accuracy, raising it from a mere 10% to 90%. This improvement significantly reduced the burden on the traffic operators by minimising false alarms but also enhanced the overall efficiency of the system. The implementation of deep learning capabilities allowed the system to analyse vast amounts of data, learning to recognise patterns and anomalies with higher precision. Moreover, the possibility for the model to continuously improve over time ensures that the AID system remains at the forefront of technological innovation, providing a robust and reliable solution for managing incidents.

Data Sources:

In this use case, data is collected from various sources, including tunnel video feeds, incident reports, and AID system logs. Each of these sources provides unique insights that contribute to the overall analysis and decision-making process.

Additional Information:

• https://www.vegvesen.no/om-oss/presse/aktuelt/2023/10/kunstig-intelligens-kan-oke-tunnelsikkerheten/
• https://kommunikasjon.ntb.no/pressemelding/18009166/kunstig-intelligens-kan-oke-tunnelsikkerheten

Point of Contact:

• pressedvvegvesen [dot] no (pressedv[at]vegvesen[dot]no)