Learning SLAM

1. Abstract

2 min read

The author condensed the abstract into the following 4 aspects:

Q1: What is Fast-LIVO2?

FAST-LIVO2 is a real-time, accurate, and robust tightly coupled multi-sensor SLAM system. It effectively integrates heterogeneous sensor data from lidar, inertial, and visual sensors using the Error-State Iterated Kalman Filter (ESIKF).
(Note: The lidar sensors include repetitive scanning multi-line rotating lidars such as Velodyne-16, XT-32, Ouster-64, Pandar128, and non-repetitive scanning solid-state lidars such as Livox-Avia. Cameras include pinhole cameras and fisheye cameras.)

Q2: What does FAST-LIVO2 solve?

  1. Heterogeneity of measurement data between LIDAR and camera;
  2. System Efficiency
  3. System accuracy
  4. System Robustness

Q3: What does FAST-LIVO2 do?

  1. Employ a sequential update strategy (first running LIO then running VIO based on the state updated by LIO) to effectively fuse heterogenous measurements from LiDAR and camera sensors into a unified system.
  2. It Directly registers raw data, avoiding complex and time consuming feature extraction processes. Measurements from both LiDAR and camera sensors are integrated into a unified voxel map (octree), enabling efficient management of 3D data.
  3. Establishes associations between LiDAR points and image patches to achieve accurate image alignment. Additionally, it dynamically refines plane parameters corresponding to LiDAR points, further enhancing alignment accuracy.
  4. Employs on demand voxel ray casting to overcome situations where data lies within LiDAR blind spots. Additionally camera eposure times are estimated online to address timestamp offsets caused by variable exposure durations under uneven illumination conditions.

Q4: What has FAST-LIVO2 demonstrated?

  1. Real-time performance (efficiency): Demonstrated through UAV navigation experiments.
  2. Accuracy and robustness: Proven by UAV-based 3D reconstruction.
  3. Scalability: Output results can be directly utilized for rendering methods like Neural Radiance Fields (NeRF) or 3D Gaussian Splatting (3D-GS).

Next section: 2. Introduction

Graph View

Backlinks