Automatic generation of 3D bridge models from 3D point clouds

• Instance extraction and filtering based purely on simple beam-bridge rules. • Adapted robust slicing algorithm with concave-hull polygon extraction that explicitly handles outliers. • Improved model for 3D semantic segmentation. • Comprehensive evaluation with public and newly acquired real bridges, with runtime analysis. Bridges play a crucial role in the transportation infrastructure and require regular inspection and maintenance to ensure safety. To enhance the efficiency of bridge maintenance, digital models, such as the Scan2BIM approach, offer a promising solution through real-time monitoring and proactive maintenance strategies. This study proposes a fully automated method of creating digital models of beam bridges from point clouds, eliminating the need for manual preprocessing or prior shape assumptions. A robust segmentation model was trained with bridge design rules being used for instantiation, followed by a slicing algorithm for model reconstruction. The method was evaluated using five publicly available bridge point clouds and three manually modeled bridges, achieving satisfactory results with a mean distance error of less than 15 cm. The automated models, which were created in an average of 2.5 minutes without human supervision, exhibited only minor discrepancies compared to manually created models, thereby significantly enhancing the efficiency of bridge modeling.