https://github.com/hku-mars/FAST-LIVO2

FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry
https://github.com/hku-mars/FAST-LIVO2

3d-reconstruction colored-point-cloud gaussian-splatting lidar-camera-fusion lidar-inertial-odometry lidar-slam mesh-reconstruction nerf sensor-fusion slam

Last synced: 12 days ago
JSON representation

FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry

• Host: GitHub
• URL: https://github.com/hku-mars/FAST-LIVO2
• Owner: hku-mars
• Created: 2024-04-22T12:49:47.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2024-08-27T05:11:01.000Z (21 days ago)
• Last Synced: 2024-08-27T06:27:27.191Z (21 days ago)
• Topics: 3d-reconstruction, colored-point-cloud, gaussian-splatting, lidar-camera-fusion, lidar-inertial-odometry, lidar-slam, mesh-reconstruction, nerf, sensor-fusion, slam
• Homepage:
• Size: 83.4 MB
• Stars: 404
• Watchers: 64
• Forks: 29
• Open Issues: 7
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# FAST-LIVO2

## FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry

### 1. Related video

Our accompanying video is now available on [**YouTube**](https://www.youtube.com/watch?v=aSAwVqR22mo&ab_channel=MARSLABHKU).

### 2. Related paper

[FAST-LIVO: Fast and Tightly-coupled Sparse-Direct LiDAR-Inertial-Visual Odometry](https://arxiv.org/pdf/2203.00893)

[FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry](https://arxiv.org/pdf/2408.14035)  

### 3. Codes & Datasets & Application

Our paper is currently undergoing peer review. The code, dataset, and application will be released once the paper is accepted.

### 4. Preview

This section showcases representative results of FAST-LIVO2 with high-resolution screenshots, allowing for easier observation of details.

#### 4.1 Online point cloud mapping results (Partial)

All sequences in FAST-LIVO2 private dataset are captured using low-cost Livox Avia LiDAR + pinhole camera.

   

     

    
"CBD Building 03" sequence (severe LiDAR and camera degeneration)



   

     

    
"Retail Street" sequence



   

     

    
"Bright Screen Wall" sequence (severe LiDAR degeneration)





    

    

    
"HIT Graffiti Wall" sequence (severe LiDAR degeneration)



   

     

    
"HKU Centennial Garden" sequence



   

     

    
"SYSU 01" sequence



   

       

     

     
Left: "Banner Wall" sequence (severe LiDAR degeneration), Right: "CBD Building 02" sequence (severe LiDAR degeneration)



        

            

      

    
Left: "HKU Landmark" sequence, Right: "HKUST Red Sculpture" sequence



    

      

    
"Mining Tunnel" sequence (severe LiDAR and camera degeneration)



   

       

     

     
"HKisland01" sequence





    

    

    
"HKairport01" Sequence (LiDAR degeneration)



#### 4.2 Mesh and texture reconstruction based on our dense colored point clouds

   

     

    
(a) and (b) are the mesh and texture mapping of “CBD

 Building 01”, respectively. (c) is the texture mapping of “Retail

 Street”, with (c1) and (c2) showing local details.



#### 4.3 Gaussian Splatting based on our dense colored point clouds

   

     

    
Comparison of ground-truth image, COLMAP+3DGS, and FAST-LIVO2+3DGS in terms of render details, computational time (time for generating point clouds and estimating poses + training time), and PSNR for a random frame in “CBD

 Building 01”.



#### 4.4 Fully Onboard Autonomous UAV Navigation

We mark a pioneering instance of employing a LiDAR-inertial-visual system for real-world autonomous UAV flights. Our UAV, equipped with LiDAR, a camera, and an inertial sensor, performs online state estimation (i.e., FAST-LIVO2), trajectory planning, and tracking control, all managed entirely by its onboard computer.



    

    
(a) shows the overall point map of the "Basement" experiment. In (a1)-(a4), white points indicate the LiDAR scan at that moment, and colored lines depict the planned trajectory. (a1) and (a4) mark areas of LiDAR degeneration. (a2) and (a3) show obstacle avoidance. (a5) and (a6) depict the camera first-person view from indoor to outdoor, highlighting large illumination variation from sudden overexposure to normal.





    

    
(a) and (b) show the enlarged point maps of the "Woods" and "Narrow Opening" experiments, respectively. The red points in (a1), (a3), and (b4) represent the LiDAR scan at that moment. (a2) and (a4) represent the first-person view at the corresponding locations. (b1)-(b3) depict the third-person view.