ITSC 2025 Paper Abstract

Paper TH-EA-T20.6

Sun, Shi (Shanghai Jiao Tong University), Li, Ruoyao (Shanghai Jiao Tong University), Wang, Yafei (Shanghai Jiao Tong University), Wang, Bowen (Shanghai Jiao Tong University), Zuo, Runheng (Shanghai Jiao Tong University), Zhang, Yichen (Shanghai Jiao Tong University)

ReUGS: Reconstruction of Unstructured Scenes Based on Geometry-Constraints Gaussian Splatting Representation

Scheduled for presentation during the Invited Session "S20b-Foundation Model-Enabled Scene Understanding, Reasoning, and Decision-Making for Autonomous Driving and ITS" (TH-EA-T20), Thursday, November 20, 2025, 14:50−15:30, Surfers Paradise 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Keywords Lidar-based Mapping and Environmental Perception for ITS Applications, Autonomous Vehicle Safety and Performance Testing, Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles

Abstract

Accurate scene reconstruction is crucial for achieving closed-loop simulation of autonomous driving. However, in unstructured scenes with sparse features, the reconstruction quality may be limited due to uneven scene information or incorrect restoration of geometric structures. To address the above issues, we propose a multi-sensor fusion reconstruction method for unstructured scenes based on 3D Gaussian splatting with multiple geometric constraints. Firstly, we designed an adaptive scene segmentation method to cope with the inherent limitation of LiDAR in initializing distant objects in outdoor environments, and initialized LiDAR Gaussians and non-LiDAR Gaussians. Subsequently, we use photometric loss, depth loss, and structural consistency loss to jointly optimize Gaussian parameters. Different density control strategies are further adopted based on Gaussian types. To retain critical information, we introduce the adaptive keyframe selection method. Finally, we evaluate our method on public datasets of different unstructured scenarios, and the results suggest that our method outperforms state-of-the-art methods in rendering quality, which is beneficial for expanding the applicability of simulation for autonomous driving.