GS-LIVO DemoPage

January 17, 2025 · View on GitHub

Project Page: gs-livo.tech (coming soon)

This repository shows the experimental results of our GS-LIVO system running on various public datasets and real-world scenarios.

System Overview and Principles

GS-LIVO (Gaussian Splatting LiDAR-Inertial-Visual Odometry) is a novel SLAM framework that seamlessly integrates LiDAR, inertial, and visual sensors. The system comprises four key modules:

Global Gaussian Map: A spatial hash-indexed octree structure that efficiently covers sparse spatial volumes while adapting to various environmental details and scales. This structure enables effective management of large-scale environments with minimal memory overhead.
Gaussian Initialization and Optimization: The system performs rapid initialization of Gaussians using both LiDAR and visual information, followed by online optimization using photometric gradients. This dual-sensor approach ensures robust and accurate scene representation.

Sliding Window Management: To maintain real-time performance, GS-LIVO employs an innovative sliding window approach for Gaussian maintenance. This includes:
- Efficient memory management between CPU and GPU
- Incremental updates to avoid redundant computations
- Strategic handling of Gaussians entering and leaving the field of view
State Estimation: The system utilizes an Iterated Extended Kalman Filter (IESKF) with sequential updates, tightly integrating LiDAR and image measurements. Unlike traditional patch-based methods, GS-LIVO achieves seamless rendering with high visual quality.

Key advantages of our approach include: