BrIMs-Based 3D Semantic Segmentation of Bridge Components Leveraging Multisensor Fusion

The high-quality semantic segmentation serves as a crucial intermediate step for the bridges management. However, achieving both high accuracy and strong generalization in segmentation remains a significant challenge. To this end, this study proposes bridge information models (BrIMs)-based three-dimensional (3D) semantic segmentation of bridge components leveraging multisensor fusion (MSF). First, an MSF-based 3D reconstruction algorithm is designed by integrating light detection and ranging, a stereo camera, and an inertial measurement unit, improving reconstruction accuracy and completeness. Second, a BrIMs-based synthetic data generation technique is designed to achieve realistic component labeling and texture mapping that closely mimic field conditions. Finally, a deep learning-based semantic segmentation network, termed RandLA-BridgeNet, is constructed to enhance segmentation generalization across diverse bridge structures. Case studies on two real bridges and an unseen cable-stayed bridge indicate that our method achieves high segmentation accuracy (overall accuracy 95.9%, mean intersection over union 96.93%) and exhibits strong generalization across various bridge types. Compared to recent works, our method delivers superior completeness, robustness, and practical value for real-world bridge inspection and digital management.