Answers are grounded in this paper and cite their sources.
Inspection of underwater bridge structures was important for ensuring bridge safety throughout the full life cycle. Three-dimensional sonar is able to reconstruct 3D models of underwater bridge structures, thereby enabling the inspection and quantification of foundation defects. However, because sonar point clouds are of low quality, manual segmentation of pile foundation point clouds remained difficult and error-prone. Therefore, a multi-view projection-based semantic segmentation method for underwater bridge sonar point clouds, termed SonarBrMvseg, was proposed in this study. First, the sonar point clouds were projected from multiple views into pseudo images, in which local geometric features were represented by inverse depth and original elevation. Subsequently, a ResUNet incorporating a coordinate attention mechanism is employed for feature extraction and pixel-wise classification. The predicted results were then reprojected into 3D labels via multi-view fusion, hence improving segmentation stability in complicated boundary regions. To alleviate class imbalance, weighted cross-entropy and Dice losses were jointly used during training. In addition, a simulated dataset was constructed for training, and underwater foundation point clouds from four bridges were collected for evaluation. The experimental results showed that the proposed method obtained a mean F1 score of 0.9288 on measured data, substantially outperforming seven comparative models, including ResNet18. Ablation analyses confirmed the effectiveness of the proposed architecture. Overall, SonarBrMvseg provides an effective solution for the analysis of underwater bridge sonar inspection data.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: