Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LVD-GS: Gaussian Splatting SLAM for Dynamic Scenes via Hierarchical Explicit-Implicit Representation Collaboration Rendering

Wenkai Zhu, Xu Li, Qimin Xu, Benwu Wang, Kun Wei, Yiming Peng, Zihang Wang

TL;DR

3D Gaussian Splatting SLAM methods struggle in dynamic outdoor scenes due to reliance on a single representation, which leads to scale drift and pose errors. LVD-GS introduces a hierarchical explicit-implicit collaboration that fuses semantic, geometric, and DINO-based cues (Sem-Geo-DINO) with an open-world and implicit residual dynamic modeling module guided by uncertainty estimates from DINO-Depth. This combination produces refined dynamic masks and multi-scale renderings, achieving state-of-the-art pose estimation and novel view synthesis on KITTI, nuScenes, and a self-collected dataset. The approach advances outdoor scene mapping robustness and lays groundwork for scalable, instance-aware cognitive navigation maps.

Abstract

3D Gaussian Splatting SLAM has emerged as a widely used technique for high-fidelity mapping in spatial intelligence. However, existing methods often rely on a single representation scheme, which limits their performance in large-scale dynamic outdoor scenes and leads to cumulative pose errors and scale ambiguity. To address these challenges, we propose \textbf{LVD-GS}, a novel LiDAR-Visual 3D Gaussian Splatting SLAM system. Motivated by the human chain-of-thought process for information seeking, we introduce a hierarchical collaborative representation module that facilitates mutual reinforcement for mapping optimization, effectively mitigating scale drift and enhancing reconstruction robustness. Furthermore, to effectively eliminate the influence of dynamic objects, we propose a joint dynamic modeling module that generates fine-grained dynamic masks by fusing open-world segmentation 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.

LVD-GS: Gaussian Splatting SLAM for Dynamic Scenes via Hierarchical Explicit-Implicit Representation Collaboration Rendering

TL;DR

3D Gaussian Splatting SLAM methods struggle in dynamic outdoor scenes due to reliance on a single representation, which leads to scale drift and pose errors. LVD-GS introduces a hierarchical explicit-implicit collaboration that fuses semantic, geometric, and DINO-based cues (Sem-Geo-DINO) with an open-world and implicit residual dynamic modeling module guided by uncertainty estimates from DINO-Depth. This combination produces refined dynamic masks and multi-scale renderings, achieving state-of-the-art pose estimation and novel view synthesis on KITTI, nuScenes, and a self-collected dataset. The approach advances outdoor scene mapping robustness and lays groundwork for scalable, instance-aware cognitive navigation maps.

Abstract

3D Gaussian Splatting SLAM has emerged as a widely used technique for high-fidelity mapping in spatial intelligence. However, existing methods often rely on a single representation scheme, which limits their performance in large-scale dynamic outdoor scenes and leads to cumulative pose errors and scale ambiguity. To address these challenges, we propose \textbf{LVD-GS}, a novel LiDAR-Visual 3D Gaussian Splatting SLAM system. Motivated by the human chain-of-thought process for information seeking, we introduce a hierarchical collaborative representation module that facilitates mutual reinforcement for mapping optimization, effectively mitigating scale drift and enhancing reconstruction robustness. Furthermore, to effectively eliminate the influence of dynamic objects, we propose a joint dynamic modeling module that generates fine-grained dynamic masks by fusing open-world segmentation 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.
Paper Structure (15 sections, 7 equations, 5 figures, 3 tables)

This paper contains 15 sections, 7 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. METHOD
  3. Hierarchical Representation Collaboration Mapping
  4. Hierarchical Representation Extraction
  5. Representation Collaboration Rendering
  6. Explicit-Implicit Joint Dynamic Modeling
  7. Uncertainty Prediction
  8. Refinement of Dynamic masks
  9. EXPERIMENTS
  10. Implementation and Experiment Setup
  11. Experiment Results
  12. Pose Estimation Results
  13. Novel View Synthesis
  14. Ablation Study
  15. CONCLUSION

Figures (5)

  • Figure 1: An overview of the chain-of-thought process, we leverage the high-level semantic understanding to construct hierarchical explicit-implicit collaborative representation constraints.
  • Figure 2: SGD-GS SLAM System Overview. A large-scale 3D Gaussian Splatting framework incorporating a multi-scale representation collaboration module, joint dynamic modeling module. We optimize camera poses using $L$ loss to establish initial pose priors, and refine these poses by incorporating 3D geometric information through scan-to-map registration follows the KISS-ICPKISS-ICP. To alleviate memory constraints, the map is partitioned into localized submaps maintained within a fixed spatial range.
  • Figure 3: Trajectory Visualization. Due to the memory constraints, other 3DGS-SLAM methods can not run to completion on all sequences, we present only our method's trajectory and error.
  • Figure 4: Novel view synthesis results on KITTI (top) , nuScenes(mid) and Self-Collected datasets (bottom). Our approach effectively handles complex dynamic environments through a Dynamic Modeling module and Representation Collaboration constraints.
  • Figure 5: Ablation study. Comparison with two novel modules: Dynamic Modeling and Representation Collaboration.