Road Surface Reconstruction using Gaussian Splatting

Read full text Summarise with VIRAT
Authors: Disa Berlin
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The areas of novel view synthesis and 3D reconstruction have recently seen major advancements with the introduction of Neural Radiance Fields and subsequently Gaussian Splatting, both of which produce accurate reconstructions of a wide variety of scenes.This thesis project explores the use of Gaussian Splatting in the area of road inspection, and aims to develop a Gaussian Splatting-based method tailored for road images paired with uncertain visual odometry poses and synthesis of bird's eye views.Two methods are proposed, where each relies on constraining the Gaussians to a surface parameterized according to a different model: a more geometrically complex ribbon model and a planar model.These methods are evaluated on twelve road sequences, and compared to the 2D Gaussian Splatting method quantitatively and qualitatively on different views.Both the ribbon model and the planar model consistently outperform the 2D Gaussian Splatting method across the tested views and metrics, suggesting that these models create more accurate road surface reconstructions from sparse data.No significant differences are observed in the accuracy of the two proposed models, suggesting that the increased geometric complexity of the ribbon model does not improve the model performance.In general, none of the proposed methods produces rendered views which are accurate enough to be used for road inspection, as road damage and road texture are often missed or inaccurately captured.This low accuracy can largely be explained by pose uncertainty, and future improvements could be made by refining the camera poses during data preprocessing or optimization.4 Results 4.1 Windscreen view . . . . .

Fields of Study & Subjects

3D Shape Modeling and Analysis Advanced Vision and Imaging 3D Surveying and Cultural Heritage

Publication details

Journal / source
Lund University Publications Student Papers (Lund University)
Publication date
1 January 2026
Language
EN
Document type
Thesis