BoltFinder: A Prototype for ML-Assisted Climbing Route Digitization

Switzerland has around 47,000 climbing routes, yet detailed information about these routes is fragmented across various platforms and printed guidebooks. The Swiss Alpine Club (SAC), a leading climbing organization, currently manages climbing sectors manually, highlighting the lack for automation. An automated solution poses challenges, such as gathering a comprehensive dataset, detecting small objects such as bolts and anchors, and using algorithmic traversal of predictions to determine accurate climbing routes for diverse areas. Tackling these challenges streamlines climbing route digitization.

BoltFinder is a proof-ofconcept solution leveraging machine learning to automate digitization of climbing routes. BoltFinder uses YOLOv11 to detect bolts and anchors in climbing wall images. Based on these detections, routes are identified through an algorithm that leverages Voronoi diagrams. A React-based frontend offers a user-friendly interface for reviewing and editing predictions and routes. The backend consists of an ExpressJS API for managing the business workflow, including the route generation algorithm, and a FastAPI responsible for handling image tiling and generating predictions. The PostgreSQL database employs a relational data model to organize and manage route structures. Model tracking is managed through MLFlow, ensuring robust performance monitoring, scalability, and adherence to MLOps principles.

The prototype uses YOLOv11 to detect bolts and anchors. Using 1024px tiles, the model exhibits a mAP\@25 score of 0.89 for bolts. YOLOv11 surpasses Faster R-CNN in precision and F1 score by ~20%. The system identifies climbing routes with an accuracy of 78% across 102 routes. 21 identified routes exhibit minor deviation whereas major deviation is observed in only one route. These results establish BoltFinder as an automated solution for climbing route digitization, providing significant value to climbers and the SAC. The prototype shows strong potential. Further improvements are possible via drone imagery, classifying different bolt types, and identifying bolt conditions like rust or wear.