SD-GCN: Saliency-based dilated graph convolution network for pavement crack extraction from 3D point clouds

Abstract

• A developed SD-GCN model for accuracy pavement crack extraction from point clouds. • Feature maps and saliency matrices construction for high-level feature encodings. • Multi-scale graph-based efficient network for inherent point feature and edge feature representations. • Cylinder-based dilated convolution strategy for computational efficiency improvement. Accurate pavement crack extraction is significant for pavement routine maintenance and potential traffic disaster minimization. Due to unordered data formats, intensity distinctions, and crack shape variations from point clouds captured by mobile laser scanning (MLS) systems, many preceding rule-based approaches and learning-based approaches cannot achieve high extraction accuracy and efficiency. To tackle these problems, we develop a saliency-based dilated graph convolution network, named SD-GCN, for pavement crack extraction from MLS point clouds. This network mainly consists of four modules. First, Module I is designed to remove off-ground point clouds. Next, two feature saliency maps are constructed to leverage both height and intensity information in Module II. Then, in Module III, the inherent point features and high-level edge features in multiple local neighborhoods are further extracted using a cylinder-based dilated convolution strategy. Finally, an MLP-based net architecture is designed for crack extraction refinement in Module IV. Experimental results exhibit that the SD-GCN model delivers an average of precision, recall, and F 1 -score of 79.5%, 77.1%, and 78.3%, respectively, which outperforms state-of-the-art methods in terms of extraction accuracy and computational efficiency.