Structure-oriented deep learning for semantic segmentation of bridge point clouds

Bridges built during the post-war construction boom are reaching the end of their design life, and their inspection still relies mainly on manual visual assessment. The specific question addressed is whether bridge components can be segmented accurately and efficiently from raw point clouds, without the manual denoising that existing methods require. This paper presents BridgeSeg, which combines a 12-dimensional structural feature embedding, a structure-aware encoder–decoder, and a loss constrained by the vertical hierarchy of bridge components, evaluated on three datasets covering 32 bridges. BridgeSeg reaches 96.6% mean intersection over union and 98.4% overall accuracy on unfiltered data, and reduces inference time by 89% compared with superpoint-based methods. This is significant for bridge owners and inspection engineers, since component-level segmentation becomes available on mainstream hardware and without preprocessing. The structure-oriented strategy can be extended to other infrastructure whose components follow known spatial rules.