RMS: Redundancy-Minimizing Point Cloud Sampling

• quick sampling of 3D LiDAR point clouds
  • pipelines using RMS are fast (low latency) and accurate
• designed for real-time LiDAR-based 6-DoF odometry/SLAM pipelines
  • both point-based (ICP-like) and feature-based (LOAM-like) methods
• single parameter only: lambda
  • depends on the SLAM pipeline (and not the environment!)
  • tuned just once given your pipeline
• deterministic (no data for learning needed)
• when is it not going to perform well (most probably):
  • the data have large orientation changes between two consecutive frames (tens of degrees)
  • under heavy noise

Paper

Published in IEEE RA-L --- pdf.

Code & How to

Installation

1. Install prerequisities (mrs_lib, PCL):

  curl https://ctu-mrs.github.io/ppa-stable/add_ppa.sh | bash
  apt-get install ros-noetic-mrs-lib ros-noetic-pcl-ros

2. Clone and build via catkin

  cd <ROS1_WORKSPACE>/src
  git clone [email protected]:ctu-mrs/RMS.git
  catkin build

How to use

1. Launch as nodelet:

roslaunch rms rms_nodelet.launch NS:=<NAMESPACE> points_in:=<POINTS IN TOPIC> points_out:=<POINTS OUT TOPIC>

2. Use as library in your code:

• add rms among dependencies in CMakeLists.txt and package.xml and include the <rms/rms.h> header file
• basic usage:

  #include <rms/rms.h>
  ...
  // Initialize
  ros::NodeHandle nh;
  mrs_lib::ParamLoader param_loader(nh, "RMS");
  RMS rms = RMS(param_loader);
  ...
  // Use
  sensor_msgs::PointCloud2::Ptr msg = ...;
  rms->sample(msg); // 'msg' now contains sampled data

• example usage: RMSNodelet implemented in src/rms_nodelet.cpp

Run it yourself

To complement the in-paper experiments, we offer comparison on the MulRan dataset by plugging its 3D LiDAR (Ouster OS1-64) data to the KISS-ICP odometry. For ROS Noetic, you may follow this workflow:

1. Click here to download the Sejong01 sequence rosbag (beware: 56 GB).
2. Install RMS (see Installation above).
3. Clone, compile, and source our KISS-ICP fork (minor changes made for ROS Noetic and launching).

cd ~/ROS1_WORKSPACE/src
git clone [email protected]:petrapa6/kiss-icp.git
cd kiss_icp
git checkout noetic
catkin build --this
source ~/ROS1_WORKSPACE/devel/setup.sh

4. Launch as:

  roslaunch kiss_icp odometry.launch bagfile:=<PATH TO ROSBAG> topic:=/mulran/velo/pointclouds use_RMS:=[true | false]

Results for the Sejong01 experiment here. APE of the experiment (voxelization in blue, RMS in orange):

How to cite

@article{petracek2024rms,
  author  = {Petracek, Pavel and Alexis, Kostas and Saska, Martin},
  title   = {{RMS: Redundancy-Minimizing Point Cloud Sampling for Real-Time Pose Estimation}},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2024},
  volume  = {9},
  number  = {6},
  pages   = {5230--5237},
  doi     = {10.1109/LRA.2024.3389820}
}

Acknowledgment

This work was supported

• by CTU grant no. SGS23/177/OHK3/3T/13,
• by the Czech Science Foundation under research project No. 23-06162M,
• by the European Union under the project Robotics and advanced industrial production (reg. no. CZ.02.01.01/00/22_008/0004590), and
• by the Research Council of Norway Award NO-321435.