Research on 3D Reconstruction Algorithm for Concrete Bridge Piers Based on Octree-GS Under the Climbing Robot Platform

Hao Du*1,2, Huifeng Wang*1,2, Zefeng Pan1,2, Yunmei Jiao1,2, Rong Gao1,2, He Huang1,2,
1 School of Electronic & Control Engineering Chang’an University, Xi’an, 710064, China, 2 Engineering Research Center of Vehicle-road Integration Perception and Intelligent Control, Universities of Shaanxi Province, Xi’an, 710064, China

TL;DR: We use the annular visual acquisition system based on a climbing robot, which collects images for training and evaluating 3DGS-based reconstruction models.

We introduce an Octree-GS-based 3D reconstruction algorithm for bridge piers that achieves SOTA performance and supports real-time rendering, providing an efficient technical solution for bridge digital inspection and the generation of digital twin foundations.



Abstract

3D reconstruction of bridge piers is critical in bridge digital inspection and digital twin technologies. In response to the limitations of traditional MVS algorithms and neural radiance fields (NeRF) in 3D reconstruction, this research proposes a novel Octree-GS-based 3D reconstruction algorithm. The algorithm learns the neural scene from multi-view images represented as 3D Gaussian distributions while integrating an octree structure with level-of-detail (LOD) techniques. It adaptively queries the corresponding LOD levels from the octree based on the observation distance and scene complexity, selecting anchor points that meet the rendering requirements. Experimental results show that the proposed 3D reconstruction algorithm achieves state-of-the-art (SOTA) performance on four bridge pier datasets, with a PSNR exceeding 32%, accurately reconstructing the bridge pier surface's geometric shape and damage distribution. Furthermore, the rendering speed exceeds 50 FPS,providing an efficient technical solution for bridge digital inspection and the generation of digital twin foundations.



Method Overview

Illustration of our proposed 3D Reconstruction Algorithm for Bridge Piers Based on Octree-GS: we start from the sparse SfM point cloud and voxelize the 3D scene of the bridge pier, using the octree structure to initialize the anchor points positions and their corresponding LOD levels. Subsequently, neural Gaussians with learnable offsets are generated for each anchor point, with their attributes (including opacity, rotation, scale, and color) dynamically predicted based on anchor point features and camera viewpoints. Next, a selection strategy is employed to choose suitable visible render anchor points, ensuring real-time stability and high rendering quality. Finally, growing and pruning operations controlled by anchor density are applied, and anchor points are adaptively refined through learnable LOD deviations to capture more details and remove outlier Gaussian points, thereby adjusting the LOD levels for each anchor point between adjacent levels.



Results

The proposed 3D reconstruction algorithm achieves SOTA performance on four bridge pier datasets, accurately reconstructing the bridge pier surface's geometric shape and damage distribution. Furthermore, it can achieve a rendering speed of 30 FPS even on a consumer-grade GPU(RTX 4060Ti), allowing engineers to inspect the surface condition of any area of the bridge pier in real-time.

PART1
PART2
PART3
PART4

Comparison with SOTA method

The proposed algorithm outperformed SOTA methods on all four bridge pier datasets. This performance improvement is primarily attributed to the following two key factors. First, we introduced anchor primitives with an octree structure, decoupling multi-scale Gaussian primitives into different LODs. Second, we proposed a novel feature bank generation method and an anchor LOD estimation approach.


Performance at Weakly Textured Regions

The proposed algorithm performs excellently in rendering background areas with weak textures on the bridge pier, allowing for a more realistic reconstruction of the bridge pier's surface textures.


Performance at Prominent Markers

The proposed algorithm excels in rendering the prominent markers on the bridge pier's surface, whether it be edge regions or small details like digits. However, its performance in reflective areas requires improvement.


Performance at Subtle Cracks

The proposed algorithm also performs excellently in reconstructing subtle cracks on the bridge pier surface, accurately restoring cracks with widths at the millimeter scale.


Performance at Evident Defects

The proposed algorithm also performs excellently in restoring evident defects on the bridge pier surface, accurately reconstructing various characteristics of the defects.


Comparison with Traditional MVS Algorithm

Compared to the traditional MVS algorithm, our method achieves clearer reconstruction results on the pier surface, particularly excelling in restoring weakly textured regions and edge details.


Visualization at different LODs

For rendering at LOD=K, the model accumulates and renders all neural Gaussians from LOD=0 to LOD=K-1. The neural Gaussians at LOD 0 are responsible for capturing the coarse features of the bridge pier surface, while higher LOD levels of neural Gaussians progressively capture finer features that were previously missed.

LOD0 LOD1 LOD2