OPENVSLAM
OpenVSLAM[1] is a open ource Visual SLAM framework. It support several types of camera models in order to achieve SLAM only based on a camera image. In agri-fields the repetitive/featuresless , texturless envoirnments leads to poor results on LIDAR based SLAM. For instance Neither with Cartographer or LIDAR SLAM package We could build a reasonable map. However openvslam perfromed quite ok in a simulated gazebo world tha covers tomato fields.
Installation and Usage of OpenVSLAM
It is best to refer to their read the docs website of OpenVSLAM here; https://openvslam.readthedocs.io/en/develop/installation.html for an updated information related to installation. we will still provide steps here to achieve installation of openvslam.
Most of openvslam dependencies will be already exstent in your Ubuntu system but make sure you have all following dependecises;
Eigen : version 3.3.0 or later.
g2o : Please use the latest release. Tested on commit ID 9b41a4e.
SuiteSparse : Required by g2o.
DBoW2 : Please use the custom version of DBoW2 released in https://github.com/shinsumicco/DBoW2.
yaml-cpp : version 0.6.0 or later.
OpenCV : version 3.3.1 or later.
Pangolin : Required for visualization and GUI.
yaml-cpp;
sudo apt-get install libyaml-cpp-dev
Some of above deps will need a source build. Which can be done as ;
DBoW2;
cd ~/
git clone https://github.com/shinsumicco/DBoW2.git
cd DBoW2
mkdir build && cd build
cmake \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX=/usr/local \
..
make -j8
sudo make install
g2o;
cd ~/
git clone https://github.com/RainerKuemmerle/g2o.git
cd g2o
git checkout 9b41a4ea5ade8e1250b9c1b279f3a9c098811b5a
mkdir build && cd build
cmake \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX=/usr/local \
-DCMAKE_CXX_FLAGS=-std=c++11 \
-DBUILD_SHARED_LIBS=ON \
-DBUILD_UNITTESTS=OFF \
-DBUILD_WITH_MARCH_NATIVE=OFF \
-DG2O_USE_CHOLMOD=OFF \
-DG2O_USE_CSPARSE=ON \
-DG2O_USE_OPENGL=OFF \
-DG2O_USE_OPENMP=ON \
..
make -j4
sudo make install
Pangolin;
Build openvslam
cd ~/
git clone https://github.com/xdspacelab/openvslam.git
cd openvslam
git checkout develop
mkdir build && cd build
cmake \
-DBUILD_WITH_MARCH_NATIVE=OFF \
-DUSE_PANGOLIN_VIEWER=ON \
-DUSE_SOCKET_PUBLISHER=OFF \
-DUSE_STACK_TRACE_LOGGER=ON \
-DBOW_FRAMEWORK=DBoW2 \
-DBUILD_TESTS=ON \
..
make -j4
sudo make install
After all of this vox_nav_openvslam should compile fine.
In order to build a map with provided server package do following;
ros2 launch vox_nav_openvslam openvslam_mapping.launch.py output_map_filename:=${HOME}/test_map.msg
Currently mono and RGBD cameras are suppoorted. RGBD is recomended and in the deafult settings we use ATM. Mono images cannot be correctly scaled to real world unlike RGBD. See the mapping.launch.py under vox_nav_openvslam and make sure the camera topics are corrctly remapped.
Note
text Visual SLAM has difficulties dealing with pure rotations. So the robot needs at least some translation as well when taking sharp turns.
Jog the robot with rqt gui plugin and visualize the map with pangolin viewer. A map with extension of .msg will be dumped to the path you passed to output_map_filename. The scripts provided in vox_nav_openvslam are able to visualize and convert this .msg to .pcd extension.
In order to visualize the created map do the following
make sure to cd ino vox_nav_openvslam/scripts .. code-block:: bash
python3 visualize_openvslam_map.py map.msg
You can convert .msg to .pcd with provided script;
python3 convert_msg_to_pcd.py map.msg out.pcd
Lastly Localization can be perfromed in a pre build map ;
where the argument is pull path to prebuild map in .msg format.