基于改进SLAM和3DGS算法的城市基础设施三维重建技术研究

 梁耀峰

摘 要

Abstract

Accurate 3D reconstruction of urban infrastructure is a key technology for driving urban digital transformation and enabling intelligent operation and maintenance. It holds significant importance for improving urban governance and ensuring the safe operation of infrastructure. However, current 3D reconstruction techniques still face bottlenecks in multi-sensor fusion localization accuracy and geometric consistency of reconstructed models, making it difficult to meet the high-precision modeling requirements of urban infrastructure scenarios. This study systematically investigates multi-sensor fusion-based localization and mapping as well as 3D reconstruction algorithms, proposing an improved algorithm framework that incorporates scene geometric priors. The proposed framework significantly enhances the overall performance of reconstructed models and contributes to the intelligent development of urban infrastructure operation and maintenance.

This thesis introduces a 3D reconstruction technology tailored for urban infrastructure scenarios, integrating LiDAR, an inertial measurement unit (IMU), and a panoramic camera. It primarily addresses two key challenges: insufficient localization accuracy in multi-sensor fusion and poor geometric consistency in 3D reconstruction. The main study contributions include: (1) Theoretical foundations of 3D reconstruction for urban infrastructure and sensor model analysis. The scene characteristics of urban infrastructure are analyzed, providing the basis for algorithmic improvements. The mathematical principles of LiDAR, IMU, and panoramic camera models are introduced, along with a detailed analysis of the baseline algorithms for simultaneous localization and mapping (SLAM) and novel view synthesis, laying the theoretical groundwork for subsequent algorithmic improvements. (2) A multi-sensor fusion localization and mapping algorithm based on an improved error-state Kalman filter (ESKF). The proposed approach includes ground point cloud segmentation and planar ground constraint improvements, constructing a joint observation model based on ESKF. It significantly reduces pose estimation drift (the z-axis localization closure error is reduced by 82.73%), and the localization accuracy surpasses other mainstream open-source algorithms. With high-precision localization results, the method generates centimeter-level accuracy point cloud maps for 3D reconstruction (achieving a mean absolute error of 1.8 cm and a mean relative error of 0.267% across multiple scenarios). (3) A novel 3D reconstruction algorithm based on an improved 3D Gaussian Splatting (3DGS) technique. By introducing LiDAR initialization, Gaussian dimensionality reduction strategies, and geometric prior-preserving optimization strategies, the proposed method effectively mitigates rendering artifacts from arbitrary viewpoints (achieving an 18% improvement in the LPIPS metric (lower is better)) and significantly enhances the geometric consistency of reconstructed urban infrastructure models.