3D Gaussian Splatting (3DGS) SLAM is widely used for high-fidelity mapping in spatial intelligence. However, current methods often rely on single-representation constraints, limiting their performance in large-scale dynamic outdoor scenes and leading to cumulative pose errors and scale ambiguity. To address these challenges, we propose LVD-GS, a novel LiDAR-Visual 3D Gaussian Splatting SLAM system for dynamic scenes. Inspired by the human coarse-to-fine comprehension process, we propose a hierarchical representations collaboration module that facilitates mutual reinforcement to optimize mapping, effectively mitigating scale drift and enhancing reconstruction robustness. Furthermore, we propose a joint dynamic modeling module that generates fine-grained dynamic masks by fusing open-world detection with implicit residual constraints, guided by uncertainty estimates from DINO-Depth features. Extensive evaluations on KITTI, nuScenes, and self-collected datasets demonstrate that our approach achieves state-of-the-art performance compared to existing methods.
Motivation of our approach. Existing 3D Gaussian Splatting SLAM methods suffer from scale ambiguity and cumulative pose errors in large-scale dynamic outdoor scenes. Our LVD-GS addresses these challenges through hierarchical explicit-implicit representation and joint dynamic modeling.
Overview of our LVD-GS framework. Our system leverages hierarchical explicit-implicit representation with LiDAR-Visual fusion for robust dynamic scene reconstruction.
Quantitative comparison on KITTI and nuScenes datasets. Our method achieves superior performance in terms of trajectory accuracy and reconstruction quality.
@article{wenkaizhu2024lvdgs,
title={LVD-GS: Gaussian Splatting SLAM for Dynamic Scenes via Hierarchical Explicit-Implicit Representation},
author={Wenkaizhu and Xu Li and Benwu Wang},
journal={arXiv preprint arXiv:2401.12345},
year={2024}
}