Speaker
Description
Accurate semantic segmentation of bridge point cloud data (PCD) is crucial for automated inspection, damage detection, and structural condition assessment of bridge infrastructure. However, existing methods often struggle to effectively capture the inherent structural relationships and spatial characteristics of bridge components. This paper presents a novel position encoding approach integrated with a hierarchical deep learning framework for bridge PCD semantic segmentation. The framework builds upon PointNet++ backbone and incorporates three key components: (1) a position encoding module that combines absolute coordinates, relative positions, and structural features, (2) an enhanced encoder with cascaded feature extraction blocks for multi-level representation learning, and (3) a progressive decoder that fuses features from multiple scales through skip connections to preserve both global context and local geometric details. Experimental results on diverse bridge datasets (from Cambridge and University of Tokyo open-source repositories) demonstrate the effectiveness of our approach, achieving 97% Overall Accuracy and 92% Mean IoU across five major bridge component categories (deck, girder, parapet, pier, and other). Cross-validation tests on bridges of varying sizes further validate the robustness and generalization capability of the proposed method. This work contributes to advancing PCD semantic segmentation by demonstrating the effectiveness of combining multi-scale spatial position encoding with hierarchical feature fusion, particularly in capturing the complex geometric relationships among bridge components.
