GS-LIVO DemoPage

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:

1. 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.

2. 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.

1. Sliding Window Management: To maintain real-time performance, GS-LIVO employs an innovative sliding window approach for Gaussian maintenance. This includes:
2. Efficient memory management between CPU and GPU
3. Incremental updates to avoid redundant computations

4. Strategic handling of Gaussians entering and leaving the field of view

5. 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: - Seamless integration of multiple sensor modalities - Efficient memory management through sliding window optimization - High-quality scene representation using Gaussian splatting - Real-time performance on both high-end GPUs and edge computing devices

Hardware & Platform Clarification

• Car Platform & Handheld Platform: Tested on Jetson Orin NX (16GB)

• Other Datasets (MARS-LVIG, Landmark, UAV, HKU): Tested on a PC with NVIDIA RTX 4090

Results on MARS-LVIG Dataset

These results were produced on PC with NVIDIA RTX 4090.

SLAM Process

SLAM Output Results

Results on Landmark Dataset

These results were produced on PC with NVIDIA RTX 4090.

SLAM Process

SLAM Output Results

Results on UAV Playground Dataset

These results were produced on PC with NVIDIA RTX 4090.

SLAM Process

SLAM Output Results

Results on FAST-LIVO HKU Dataset

These results were produced on PC with NVIDIA RTX 4090.

SLAM Process

SLAM Output Results

Vehicle Implementation

Tested on Jetson Orin NX (16GB).

Implementation of GS-LIVO on a real vehicle with A* LQR path planning:

Edge Computing Deployment

Tested on Jetson Orin NX (16GB).

Real-time deployment on Jetson Orin NX 16GB: