Hybrid Workflow for Digital Twin Creation: Human-in-the-Loop Segmentation and Geometry-Aware Modeling of Bridge Structures

Abstract

Reliable as-is modeling of existing infrastructure is a cornerstone of structural health monitoring and digital twin creation. While data-driven segmentation approaches have advanced automation for common bridge types, they often struggle to generalize to atypical structures. This paper instead proposes an interactive segmentation and modeling technique that aims to reduce the time needed for producing accurate geometric models and segmentation masks by considering both fine local features as well as large-scale geometry. Our method begins with a graph-based oversegmentation of the input point cloud, compressing millions of points into a superpoint graph optimized for real-time interaction. On this graph, we formulate user-guided segmentation as an energy minimization problem, incorporating both local features and region-growing cues to propagate sparse user input. Once segmented, each component undergoes geometry-specific processing: planar elements are modeled via RANSAC-based decomposition and intersection analysis, while irregular surfaces are reconstructed using adaptive triangulation informed by local point density.The resulting surface representations are converted into IFC-compliant building elements, enabling integration with existing BIM systems. We evaluate the method on the Nibelungen Bridge in Worms, Germany, achieving over 99% segmentation accuracy with minimal user input and a final geometric RMSE of 5.8 cm.