SuMa++: Efficient LiDAR-based Semantic SLAM

This repository contains the implementation of SuMa++, which generates semantic maps only using three-dimensional laser range scans.

Developed by Xieyuanli Chen and Jens Behley.

SuMa++ is built upon SuMa and RangeNet++. For more details, we refer to the original project websites SuMa and RangeNet++.

An example of using SuMa++: ptcl

Table of Contents

  1. Introduction
  2. Publication
  3. Dependencies
  4. Build
  5. How to run
  6. More Related Work
  7. License

Publication

If you use our implementation in your academic work, please cite the corresponding paper:

@inproceedings{chen2019iros, 
		author = {X. Chen and A. Milioto and E. Palazzolo and P. Giguère and J. Behley and C. Stachniss},
		title  = {{SuMa++: Efficient LiDAR-based Semantic SLAM}},
		booktitle = {Proceedings of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS)},
		year = {2019},
		codeurl = {https://github.com/PRBonn/semantic_suma/},
		videourl = {https://youtu.be/uo3ZuLuFAzk},
}

Dependencies

  • catkin
  • Qt5 >= 5.2.1
  • OpenGL >= 4.0
  • libEigen >= 3.2
  • gtsam >= 4.0 (tested with 4.0.0-alpha2)

In Ubuntu 16.04: Installing all dependencies should be accomplished by

sudo apt-get install build-essential cmake libgtest-dev libeigen3-dev libboost-all-dev qtbase5-dev libglew-dev libqt5libqgtk2 catkin

Additionally, make sure you have catkin-tools and the fetch verb installed:

sudo apt install python-pip
sudo pip install catkin_tools catkin_tools_fetch empy

Build

rangenet_lib

To use SuMa++, you need to first build the rangenet_lib with the TensorRT and C++ interface. For more details about building and using rangenet_lib you could find in rangenet_lib.

SuMa++

Clone the repository in the src directory of the same catkin workspace where you built the rangenet_lib:

git clone https://github.com/PRBonn/semantic_suma.git

Download the additional dependencies (or clone glow into your catkin workspace src yourself):

catkin deps fetch

For the first setup of your workspace containing this project, you need:

catkin build --save-config -i --cmake-args -DCMAKE_BUILD_TYPE=Release -DOPENGL_VERSION=430 -DENABLE_NVIDIA_EXT=YES

Where you have to set OPENGL_VERSION to the supported OpenGL core profile version of your system, which you can query as follows:

$ glxinfo | grep "version"
server glx version string: 1.4
client glx version string: 1.4
GLX version: 1.4
OpenGL core profile version string: 4.3.0 NVIDIA 367.44
OpenGL core profile shading language version string: 4.30 NVIDIA [...]
OpenGL version string: 4.5.0 NVIDIA 367.44
OpenGL shading language version string: 4.50 NVIDIA

Here the line OpenGL core profile version string: 4.3.0 NVIDIA 367.44 is important and therefore you should use -DOPENGL_VERSION = 430. If you are unsure you can also leave it on the default version 330, which should be supported by all OpenGL-capable devices.

If you have a NVIDIA device, like a Geforce or Quadro graphics card, you should also activate the NVIDIA extensions using -DENABLE_NVIDIA_EXT=YES for info about the current GPU memory usage of the program.

After this setup steps, you can build with catkin build, since the configuration has been saved to your current Catkin profile (therefore, --save-config was needed).

Now the project root directory (e.g. ~/catkin_ws/src/semantic_suma) should contain a bin directory containing the visualizer.

How to run

Important Notice

  • Before running SuMa++, you need to first build the rangenet_lib and download the pretrained model.
  • You need to specify the model path in the configuration file in the config/ folder.
  • For the first time using, rangenet_lib will take several minutes to build a .trt model for SuMa++.
  • SuMa++ now can only work with KITTI dataset, since the semantic segmentation may not generalize well in other environments.
  • To use SuMa++ with your own dataset, you may finetune or retrain the semantic segmentation network.

All binaries are copied to the bin directory of the source folder of the project. Thus,

  1. run visualizer in the bin directory by ./visualizer,
  2. open a Velodyne directory from the KITTI Visual Odometry Benchmark and select a “.bin” file,
  3. start the processing of the scans via the “play button” in the GUI.

More Related Work

OverlapNet – Loop Closing for 3D LiDAR-based SLAM

This repo contains the code for our RSS2020 paper: OverlapNet – Loop Closing for 3D LiDAR-based SLAM.

OverlapNet is a modified Siamese Network that predicts the overlap and relative yaw angle of a pair of range images generated by 3D LiDAR scans, which can be used for place recognition and loop closing.

This repo contains the code for our IROS2020 paper: Learning an Overlap-based Observation Model for 3D LiDAR Localization.

It uses the OverlapNet to train an observation model for Monte Carlo Localization and achieves global localization with 3D LiDAR scans.

License

Copyright 2019, Xieyuanli Chen, Jens Behley, Cyrill Stachniss, Photogrammetry and Robotics Lab, University of Bonn.

This project is free software made available under the MIT License. For details see the LICENSE file.

GitHub

View Github