SegMap is a map representation based on 3D segments allowing for robot localization, environment reconstruction, and semantics extraction. The SegMap code is open-source (BSD License) and has been tested under Ubuntu 14.04, 16.04 and ROS Indigo, Kinetic. Please remember that this is on-going research code which is subject to changes in the future.
R. Dubé, A. Cramariuc, D. Dugas, H. Sommer, M. Dymczyk, J. Nieto, R. Siegwart, and C. Cadena. SegMap: Segment-based mapping and localization using data-driven descriptors The International Journal of Robotics Research (IJRR), 2019 pdf
R. Dubé, A. Cramariuc, D. Dugas, J. Nieto, R. Siegwart, and C. Cadena. "SegMap: 3D Segment Mapping using Data-Driven Descriptors." Robotics: Science and Systems (RSS), 2018. pdf - video
R. Dubé, MG. Gollub, H. Sommer, I. Gilitschenski, R. Siegwart, C. Cadena and , J. Nieto. "Incremental Segment-Based Localization in 3D Point Clouds." IEEE Robotics and Automation Letters, 2018. pdf
R. Dubé, D. Dugas, E. Stumm, J. Nieto, R. Siegwart, and C. Cadena. "SegMatch: Segment Based Place Recognition in 3D Point Clouds." IEEE International Conference on Robotics and Automation, 2017. pdf - video
- 3D CNN encoder – decoder
- ICP based LiDAR odometry
- Dynamic voxel grid
- Single and multi-robot SLAM back-end
- Generic incremental region-growing segmentation
- Incremental geometric verification
The SegMap repository contains the following modules:
- segmap: The C++ library for 3D segment mapping.
- segmap_ros: ROS interface for segmap.
- segmapper: Example application using segmap and ros.
- laser_slam: Backend for the example application, based on LiDAR sensor data.
- segmappy: Python library for training and evaluating the neural network models.
This section provides a step by step guide to installing the full flavor of SegMap. Advanced: it is also possible to use the SegMap C++ library standalone in a C++ project, or integrate the C++ library and ROS interface in a ROS project.
First install the required system packages:
$ sudo apt-get install python-wstool doxygen python3-pip python3-dev python-virtualenv dh-autoreconf
Set up the workspace configuration:
$ mkdir -p ~/segmap_ws/src
$ cd ~/segmap_ws
$ catkin init
$ catkin config --merge-devel
$ catkin config --cmake-args -DCMAKE_BUILD_TYPE=Release
Then use wstool for fetching catkin dependencies:
$ cd src
$ git clone https://github.com/ethz-asl/segmap.git
$ wstool init
$ wstool merge segmap/dependencies.rosinstall
$ wstool update
If you do not plan on using the deep learned descriptors in SegMap you should at least install the CPU version of Tensorflow to compile the whole package. Otherwise we recommend the GPU version. |
---|
The precompiled version of Tensorflow can only be used with Ubuntu 14.04, for further explanations see here (requirements for the GPU version here). For Ubuntu 16.04 and 18.04 you must compile tensorflow from source. After compiling tensorflow, generate the pip package and continue with these instructions, installing the custom generated package instead of the precompiled one. |
---|
SegMap relies on the package tensorflow_ros_cpp for linking to the tensorflow C++ API. All our code has been tested with Tensorflow 1.8.
$ virtualenv ~/segmappyenv
$ source ~/segmappyenv/bin/activate
(segmappyenv)$ pip install --upgrade pip
(segmappyenv)$ pip install catkin_pkg empy pyyaml
(segmappyenv)$ pip install tensorflow-gpu==1.8.0
For more details on the optional compile flags or in case of issues compiling see the FAQ.
$ cd ~/segmap_ws
$ catkin build tensorflow_ros_cpp
Finally, build the segmapper package which will compile all dependencies and SegMap modules:
$ cd ~/segmap_ws
$ catkin build segmapper
Installing segmappy allows you to train data-driven models yourself to use with SegMap.
$ cd src/segmap/segmappy/
$ source ~/segmappyenv/bin/activate
(segmappyenv)$ pip install .
Make sure to source the SegMap workspace before running the segmapper demonstrations:
$ source ~/segmap_ws/devel/setup.bash
To train new models see intructions here.
To download all necessary files, copy the content of the segmap_data into ~/.segmap/
. If you installed the segmappy python package you can run the automated download script. Note: These models have been trained using Tensorflow 1.8 and are only guaranteed to work for that version.
An online SLAM example can be run with
$ roslaunch segmapper kitti_loop_closure.launch
A localization example can be run with
$ roslaunch segmapper kitti_localization.launch
An online SLAM example with data-driven descriptor can be run with
$ roslaunch segmapper cnn_kitti_loop_closure.launch
You can now visualize the reconstructed target map in rviz by subscribing to /segmatch/target_reconstruction
.
More details on the demonstrations can be found here.
SegMap is released under BSD 3-Clause License
Thank you for citing the related publication if you use SegMap in academic work:
@article{doi:10.1177/0278364919863090,
title ={SegMap: Segment-based mapping and localization using data-driven descriptors},
author = {Renaud Dubé and Andrei Cramariuc and Daniel Dugas and Hannes Sommer and Marcin Dymczyk and Juan Nieto and Roland Siegwart and Cesar Cadena},
journal = {The International Journal of Robotics Research},
volume = {39},
number = {2-3},
pages = {339-355},
year = {2020},
doi = {10.1177/0278364919863090}
}
@inproceedings{segmap2018,
title={{SegMap}: 3D Segment Mapping using Data-Driven Descriptors},
author={Dub{\'e}, Renaud and Cramariuc, Andrei and Dugas, Daniel and Nieto, Juan and Siegwart, Roland and Cadena, Cesar},
booktitle={Robotics: Science and Systems (RSS)},
year={2018}
}
@inproceedings{segmatch2017,
title={SegMatch: Segment based place recognition in 3d point clouds},
author={Dub{\'e}, Renaud and Dugas, Daniel and Stumm, Elena and Nieto, Juan and Siegwart, Roland and Cadena, Cesar},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
pages={5266--5272},
year={2017},
organization={IEEE}
}
If you use the incremental version for the segmentation or geometric verification algorithms (used in the provided configuration files), please consider citing the describing paper:
@article{incremental2018,
title={Incremental Segment-Based Localization in {3D} Point Clouds},
author={Dub{\'e}, Renaud and Gollub, Mattia G and Sommer, Hannes and Gilitschenski, Igor and Siegwart, Roland and Cadena, Cesar and Nieto, Juan},
journal={IEEE Robotics and Automation Letters},
volume={3},
number={3},
pages={1832--1839},
year={2018},
publisher={IEEE}
}
We would be very grateful if you would contribute to the code base by reporting bugs, leaving comments and proposing new features through issues and pull requests. Please see the dedicated wiki page on this topic and feel free to get in touch at renaud.dube(at)sevensense(dot)ch, dugasd(at)ethz(dot)ch and crandrei(at)ethz(dot)ch. Thank you!