Segmentation of large-scale masonry arch bridge point clouds with a synthetic simulator and the BridgeNet neural network

Abstract

Masonry arch bridges constitute the majority of the European bridge stock and feature a wide range of geometric characteristics. Due to a general lack of construction drawings, their geometry is difficult to parameterize. Laser scanning devices are commonly used to capture bridge geometry. However, this requires time-consuming segmentation of point clouds into their constituent components to extract key geometric parameters. To increase efficiency, a 3D deep learning neural network called BridgeNet is proposed that can automate the segmentation. To tackle the scarcity of labelled point cloud data, a synthetic dataset is created to train BridgeNet. BridgeNet is subsequently tested on real point clouds and achieves state-of-art performance, demonstrating the utility of synthetic training data and the advantages of the new network design. Segmented components are then fitted with primitive shapes by using Random Sample Consensus based algorithms to characterize key geometric parameters to assist assessments and inspections.