Deep-Learning-Based Structure-Aware 3D Reconstruction For Cable-Stayed Bridges

With the development of new technologies of sensors and computer science,structural health monitoring systems have been widely used in the field of civil engineering.Based on computer vision,smart detection technology solves many problems that exist in traditional manual visual inspection,including the cost,efficiency,quantifying,and subjectivity problems.However,current visual detection technology still suffers from the limitation of globality due to the lack of structure of image data,structure information of a single component,and the application ability to overall evaluation.Therefore,the combination of image data-based detection and structured 3D reconstruction model becomes a trend of smart vision detection.Traditional structure-aware 3D reconstruction methods are usually limited to specific structure types,such as concrete girder bridges,and are short of robustness to point cloud quality.The research on structure-aware 3D reconstruction of cable-stayed bridge systems is still in a fledging period.Therefore,this paper studies the 3D reconstruction of cable-stayed bridges based on computer vision and deep learning.The main research contents include:The structural perception modeling method of cable-stayed bridges considering structural relations is studied.Firstly,a hierarchical binary tree model is proposed to describe the high-level relationship of the cable-stayed bridge structure,namely the topological relationship and similarity relationship among structural components.Secondly,based on the leaf nodes of the binary tree,the geometrical model and the voxel model are adopted to model the low-level 3D shapes.Therefore,the 3D model of an entire cable-stayed bridge is embedded as a binary tree model.Besides,the proposed binary tree model is robust to noise and partial scans in point clouds.The deep learning method of cable-stayed bridge structure-aware 3D reconstruction is studied.In the feature extraction part,a multi-view convolutional neural network and a point cloud neural network are used to extract hybrid features from images and 3D point clouds.The decoding part of the model uses the recursive binary tree network to construct the high-level structure graph layouts and predict the low-level geometry of 3D shapes.The effectiveness of the proposed method is verified.Two actual cable-stayed bridges are employed as examples to evaluate the proposed method.Test results demonstrate that the proposed method successfully reconstructs the bridge model with structural components and their relations.Quantitative results indicate that the predicted models achieved an average₁ score of 99.01%and a mesh-to-cloud distance of 1.78 m.The achieved result is similar to that obtained using the manual reconstruction approach in terms of component-wise accuracy,and it is considerably better than that obtained using the manual approach in terms of spatial accuracy.