| description abstract | Deep learning-based scan-to-building information models (BIMs) approaches have gained popularity in generating as-built BIMs for highway bridges. However, several critical challenges emerge. First, the acquisition of large-scale training data is prohibitively expensive. Second, the complex geometry of bridges, such as variable curvature and cross sections, poses difficulties for modeling. Lastly, these generated models often lack a parametric definition. To address these challenges, this paper introduces an enhanced scan-to-BIM framework that uses low-cost synthetic point clouds and parametric modeling. This approach contains two main components: (1) semantic segmentation using augmented datasets, and (2) a projection-based parametric modeling method. Through rigorous experiments, it is evident that augmenting the training process with synthetic data significantly improves model performance, yielding up to a 12.2% segmentation improvement in this work. In terms of modeling, the reconstructed model showed a marginal mean difference of 0.06 m against the ground truth. Notably, when applied to real-world bridges, the framework demonstrated a comparable accuracy level, with deviations primarily stemming from occlusions in the bridge abutments. In conclusion, this study highlights the effectiveness of the proposed framework in creating as-built BIM models for highway bridges. Additionally, it emphasizes the significance of combining synthetic and real data for optimal accuracy and proves its potential in modern highway infrastructure applications. | |
