Voxelized 3D Gaussian Representation for Dense Visual SLAM on Embedded Vision System

Abstract

Recent work has shown that 3D Gaussian-based SLAM enables high-quality reconstruction, accurate pose estimation, and real-time rendering of scenes. However, when we revisit the existing GS-based SLAM systems, we observe that the redundancy of the Gaussian ellipsoids created by SLAM systems is significantly superior to those of the original 3D Gaussian Splatting methods. We further quantitatively demonstrate the geometric similarity and the redundancy of 3D Gaussian scene representation, leading to high memory and storage costs and limiting its broader application in embedded platforms. To address this limitation, we propose a voxelized compact 3D Gaussian representation for SLAM system that reduces the number and the parameter size of Gaussian ellipsoids. A novel voxel-anchored 3D Gaussian scene representation method is proposed to achieve compact and accurate scene reconstruction. A sliding window-based online masking strategy is first proposed to reduce the redundant ellipsoids. Then, a novel residual codebook-based quantization method is proposed to further compress 3D Gaussian attributes. Robust and accurate pose estimation is achieved by a local-to-global bundle adjustment method with icp loss. Extensive experiments demonstrate that our method achieves faster training, rendering speed (226%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document} increase), and low memory usage (2.21×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} compression) while maintaining the state-of-the-art (SOTA) quality of the scene representation. Moreover, we also test our framework on an embedded platform, Jetson Xavier MIC-730AI, showing the potential of application in embedded system. We also collect a Neural SLAM dataset using our mobile robot’s embedded hardware platform, incorporating data from multiple sensor modalities to further validate the effectiveness of our algorithm on a real-world system across various indoor and outdoor scenes. We will release our code in https://github.com/dtc111111/VCGS-SLAM.