Case study Smart Tunnel
To improve the accuracy of its Automatic Incident Detection (AID) system, ISSD switched to a deep learning approach. fortiss supported this transition through technical coaching and helped ISSD to analyse and optimise the robustness of the neural networks against real-world interference factors. The result is a solution that halves false alarms, increases detection speed and, at the same time, reduces hardware costs per video stream.
Challenge
Conventional AID systems based on traditional computer vision are increasingly reaching their limits. Tunnel operators often face cognitive overload as they have to monitor dozens of camera feeds simultaneously, whilst the system frequently triggers false alarms. ISSD wanted to integrate deep learning to improve detection accuracy. The challenge was to ensure robustness, interpretability and cost-efficiency in a productive, safety-critical environment. Factors such as adverse weather conditions, image noise or changes in lighting had to be reliably processed without significantly increasing hardware requirements.
Solution
fortiss provided in-depth technical coaching in the field of neural networks. The team helped ISSD to analyse the decision-making processes of the object recognition model and to assess its reliability under challenging conditions (e.g. image artefacts or sensor noise). Together, they scrutinised performance evaluations and developed targeted model improvements to further enhance robustness.
The final solution was implemented on the basis of the Intel Myriad-X VPU in a highly scalable hardware configuration. This enabled highly efficient real-time video processing directly ‘at the edge’, which significantly reduced the cost per stream.
Result
- Significant reduction in errors: A reduction in the false alarm rate of around 50% compared to the previous solution, whilst maintaining complete event capture
- Improved performance: A significant improvement in detection times for safety-critical incidents
- Cost-efficiency: A reduction in hardware costs per video stream, despite increased computing requirements, through optimised VPU utilisation
- Validated robustness: Successful test run of the improved engine in a real-world tunnel monitoring environment
- Knowledge transfer: Development of in-house expertise at ISSD for analysing and optimising the interpretability of AI models.
Outcome
A deep learning-based AID system can significantly outperform conventional approaches in terms of both accuracy and cost-effectiveness. Thanks to targeted support from fortiss, ISSD was able to massively increase the reliability of its AI model under real-world operating conditions. The project highlights just how crucial close cooperation between research and industry is for the scalable deployment of AI in safety-critical areas of infrastructure.
Project partner
