Automated 3D Reconstruction of Bridges Based on 3D Gaussian Splatting

Digital twin and metaverse play a crucial role in realizing the intelligent operation and maintenance of bridges, in which the automated 3D reconstruction of bridges is the key to build the digital twin model and metaverse platform. To address the challenge of integrating geometric and semantic reconstruction in existing 3D bridge reconstruction methods, this study proposed a high-fidelity automated 3D bridge reconstruction method based on 3D Gaussian splatting (3DGS). First, to capture the boundary features of bridge components, a semantic segmentation model for bridges was proposed using the Segment Anything Model (SAM), enabling high-precision semantic segmentation of multi-view UAV-captured bridge images and generating semantic masks for components such as decks and bridge towers. Next, semantic feature attributes were incorporated into the 3D Gaussian kernel functions of scene representation, achieving adaptive characterization of complex geometric and semantic features. Finally, 3D Gaussians were projected into 2D image space, and differentiable rasterization algorithms were utilized to generate semantic feature maps and rendered images. Through joint optimization of geometric, appearance, and semantic attributes, a unified 3D reconstruction was realized, balancing geometric fidelity and semantic accuracy. In this paper, Nansha Bridge in Guangdong Province is used as an example for validation, and the experimental results show that the reconstructed 3D bridge model achieves high-fidelity restoration of texture details and semantic information, with a mean Intersection over Union (mIoU) of 87. 2% and an overall accuracy of 91.6%. The proposed method effectively resolves geometric detail loss and semantic fragmentation in traditional approaches, providing a 3D foundational model with both physical precision and semantic interpretability for lifecycle digital twins of bridges.