better_fastlio2

To be a better FAST-LIO2

0 Introdunction

• Front_end: FAST-LIO2 + Dynamic Removal + Yolo(Optional)
• Back_end: Scan Context + GPS(Optional) + GTSAM
• Application: Joint Pose-graph Optimization using iSAM2, Fast and Robust ICP relocalization

Note

• The offline dynamic removal module using segmented curved-voxel occupancy descriptor (T-GRS 2024) can be found in repo
• The modified multi-session mapping module based on LT-Mapping (ICRA 2022) can be found in repo
• The new version of online relocalization module name LiLoc (ICRA 2025) can be found in repo
• Some details of sensor platform and application can be found in repo

1 Prerequisites

• Ubuntu 20.04 and ROS Noetic
• PCL >= 1.10 (default for Ubuntu 20.04)
• Eigen >= 3.3.4 (default for Ubuntu 20.04)
• GTSAM >= 4.0.3 (test on 4.2(a))
• Livox_ros_driver
• Darknet_ros

3 Build

3.1 Source Code

cd YOUR_WORKSPACE/src
git clone https://github.com/Yixin-F/better_fastlio2
cd ..
catkin_make

3.2 Docker

…

One problem: if you want to use RViz or Gazebo, exit the container and enter again

docker start better-fast-lio2
docker exec -it better-fast-lio2 /bin/bash

4 How to Use

4.1 LIO Mapping using Livox, Velodyne, Ouster or Robosense

In this section, we developed a more comprehensive FAST-LIO2 version including dynamic removal using SCV-OD (accepted by T-GRS) and YOLO(optional), optimization backend using Scan Context, GPS(optional) and GTSAM. At the same time, we rewritted the mechanism of i-kdtree reconstruction to be suitable for low-power embedded devices. The basic framework is illustrated in the following figure.

You can run it by the following commands.

source ./devel/setup.bash
roslaunch fast_lio_sam mapping_*.launch

For other application, you need to first check the Config/*.yaml about the settings for different LiDAR types, we list parameters here. "-" means that it depends on your own project.

Note that if you wanna use the dynamic removal module, you have to make the "src/laserMapping.cpp line 2271-2307" effect. But differnet from the origianl dynamic removal method in T-GRS paper, we utilized the voxel center to align consecutive keyframes, which improves the real-time performance but also sacrifices some precision. So, we do not recommend using the Velodyne VLP16 for testing because its scan point cloud is too sparse. This dynamic removal module can be showed by the following figure.

1. Here we list some important functions in src/laserMapping.cpp as follows:

1. Here we list some important functions in include/dynamic-remove/tgrs.cpp as follows:

1. Here we list some important functions in include/sc-relo/Scancontext.cpp as follows:

After you have run the command, there are several files being generated in the filefold "rootDir/*" as follows:

We show some simple results:

Now, we give some suggestion to improve this special LIO framework: Firstly, change the point-cloud map management mechanism. You can use the sparse method like hash table similar to Faster-LIO.

Secondly, improve the residual calculation method. The point-to-plane registrition during LiDAR measurement, which utilizes generalized patches instead of actual planes may lead to inconsistencies in registration. You can use true planes modeling by uncertainty.

Thirdly, find a better loop closure descript