Extraction of key geometric parameters from segmented masonry arch bridge point clouds

Abstract

Masonry arch bridges constitute the majority of the European bridge stock. Most of these bridges were constructed in the 19th century and feature a wide range of geometric characteristics. Since construction drawings rarely exist, the first step in the assessment of these bridges is the characterisation of their in-situ geometry, which may involve significant geometric distortions. In recent years, LIDAR devices have been widely used by bridge owners due to their ability to remotely and rapidly collect point cloud data. To enable the engineering assessment practice to benefit from this data, this research uses the recently developed deep learning (DL) neural network BridgeNet to autonomously segment masonry bridge point clouds into different components. Due to the limited availability of 3D point clouds, BridgeNet is trained using a synthetic multi-span masonry arch bridge dataset; the network is then tested on real arch bridge point clouds. By fitting appropriate primitive shapes to bridge component point clouds using Random Consensus Sampling (RANSAC) techniques the bridge geometry is effectively characterised by a few parameters.