Speaker
Description
The High-Luminosity Large Hadron Collider (HL-LHC) project introduces increased radiation levels in the Long Straight Sections (LSS),
creating a strong demand for automated measurement and remote alignment solutions.
To address this challenge, CERN has developed the Full Remote Alignment System (FRAS) for the LSS, together with Ecartometry Measurement by Automatic Photogrammetric Survey (EMAPS).
EMAPS enables the transfer of the geometric reference frame from the FRAS to the tunnel wire-based measurement system,
ensuring sufficient overlap between the two systems and maintaining alignment continuity.
This contribution investigates the application of neural networks to enhance the existing close-range photogrammetric processing pipeline used for accelerator alignment.
The work focuses on the detection and measurement of photogrammetric targets and stretched wires.
A YOLO26s object-detection model was trained on a combined dataset of synthetic and real images for photogrammetric target detection.
Target-centre localisation was achieved using a modified ElDet model trained on real images annotated with a commercial photogrammetry software package.
For stretched-wire detection, a YOLO26n-seg segmentation model was trained on real images using annotations derived from filtered outputs of the existing processing workflow.
The proposed solution was applied to wire offset measurements acquired by EMAPS during LHC technical stop periods.
Deep learning-based models demonstrated promising performance relative to conventional analytical approaches, achieving sub-20 µm precision.
In particular, the proposed method significantly improved the reliability of detecting photogrammetric targets and thin stretched wires in high-resolution images,
reducing both missed detections and false positives by approximately 80%.
These results constitute a significant step toward real-time, fully automated photogrammetric processing for accelerator alignment applications
while maintaining the required measurement precision.
