A vision-based abnormality detection system in the LHC tunnel

Abstract

CERN, the European Organisation for Nuclear Research has more than 50 km of tunnels hosting machinery in extreme environments. One of these tunnels hosts the Large Hadron Collider (LHC), the world's largest particle accelerator. To ensure safety, the 27km long tunnel structure, which lies at around 100 metres below the ground, together with the equipment within it, need to be regularly monitored. This raised the need for a remotely operated surveying system and consequently, a Train Inspection Monorail (TIM) was installed in the LHC tunnel. The TIM gathers visual data as the train slithers along the tunnel using a camera fixed on an arm extending downwards from one of its wagons. However, up till now, the images captured were only used for data record purposes. This work focused on developing a vision-based inspection system that gathers data from small and low-cost camera equipment placed on the TIM. Using simple, yet effective image processing techniques, a system with a robust algorithm that is able to monitor changes on the LHC tunnel linings was devised. Images shot in tunnel environments are characterised by low light. A shading correction algorithm was thus designed to enhance the image brightness and eliminates shadows. The difference in capture location over time demanded the need for position offset correction before image comparison. This was achieved using image mosaicing, where, unlike current solutions, the proposed method utilises binary edges as high-level instead of low-level features, as the latter are either not present or rare in tunnel environments. Once the two closest reference images to the current image are stitched, the Survey Mapper locates the survey image in the mosaic image. Finally, the Change Detector uses image differencing, binary image comparison and optical flow analysis to detect any changes between one inspection and another. The system achieves a high sensitivity of 83.5% and an 82.8% precision, as well as an average accuracy of 81.4%. This provides a reliable tunnel wall change monitoring framework which is able to detect changes as small as around 10 cm in any dimension. The proposed system is also configurable through different parameters to adapt to the scenario in place, making it useable in other tunnel environments and not exclusive to the LHC tunnel.