This comparison is part of a localization robotic course at the Universidade Federal da Bahia (UFBA) master's degree program. The comparison here is intended to minimally mimic the excellent work made by Alexey Merzlyakov and Steve Macenski, who wrote the paper A Comparison of Modern General-Purpose Visual SLAM Approaches. It is worth highlighting that this work is not intended to replace the work made by these authors.
The comparison has changed slightly from the original work. This work compared the methods ORB-SLAM3, OpenVSLAM and Hector SLAM using data gathered in a Gazebo Simulation.
The Dockerfile provides the dependencies to execute the Visual SLAM methods of the comparison and the installation steps to configure the container are described in the section Configuring the docker environment. However, if you want proceed without the docker container, the commands listed in the Dockerfile and entrypoint.sh are a good start points to configure the environment in your machine.
The visual SLAM methods were initially tested with two open datasets available: TUM RGB-D and EuroC MAV. However the comparison was executed using ROS bags recorded using the LAR Gazebo package, which reproduce the Robotics Laboratory at UFBA. The bags were recorded using the LiDAR Sick LMS1XX and the RGB-D Sensor Realsense D315 assemble on the robot Husky of ClearPath. The bags were generated during three trajectories: Rectangular Trajectory, Figure-Eight Trajectory and, Straight Line Trajectory, sorted ascending by the complexity for the VSLAM methods. More info about the bags in ROS Bags Content.
First of all, install the Docker Engine on your computer:
- If don't have NVIDIA GPU, follow the official instructions in Install Docker Engine on Ubuntu
- Otherwise, follow the instructions on the NVIDIA document's page
After the Docker Engine installation, clone this repository into your machine
git clone https://github.com/mateusmenezes95/vslam_comparison.git
As you see, it was created a Dockerfile and bash scripts to ease the replication of the comparison through a simulation using ROS Noetic and Gazebo 11. To set up the docker container, follow the next instructions.
Go into the folder docker containing bash scripts and the Dockerfile
cd docker/
Now run, which will build a docker image of ROS Noetic
./build_image.sh
After the build phase, just run the container
./run-container.sh
To see if is everything ok, in the terminal launched, run
roscore
Execute the ./run-container.sh in a new terminal to launch a new terminal inside the same container launched a priori. Then, run the listener node. Nothing will happen until you launch the talker node
rosrun roscpp_tutorials listener
To launch the talker node, again in a new terminal repeating the execution of ./run-container.sh, run
rosrun roscpp_tutorials talker
You must see something like
- From the
talkernode:
mateus_docker@dell:~$ rosrun roscpp_tutorials talker
[ INFO] [1664851703.004928103]: hello world 0
[ INFO] [1664851703.105062777]: hello world 1
[ INFO] [1664851703.205134225]: hello world 2
[ INFO] [1664851703.305133228]: hello world 3
[ INFO] [1664851703.405134370]: hello world 4
[ INFO] [1664851703.505133031]: hello world 5
[ INFO] [1664851703.605162346]: hello world 6
[ INFO] [1664851703.705122720]: hello world 7
[ INFO] [1664851703.805129064]: hello world 8
[ INFO] [1664851703.905136324]: hello world 9
[ INFO] [1664851704.005136090]: hello world 10
...
- From the
listenernode:
mateus_docker@dell:~$ rosrun roscpp_tutorials listener
[ INFO] [1664851703.305901909]: I heard: [hello world 3]
[ INFO] [1664851703.405803441]: I heard: [hello world 4]
[ INFO] [1664851703.505634738]: I heard: [hello world 5]
[ INFO] [1664851703.605727613]: I heard: [hello world 6]
[ INFO] [1664851703.705598423]: I heard: [hello world 7]
[ INFO] [1664851703.805713104]: I heard: [hello world 8]
[ INFO] [1664851703.905704093]: I heard: [hello world 9]
[ INFO] [1664851704.005709959]: I heard: [hello world 10]
...
To kill the nodes press ctrl+c to the talker, listener, and roscore terminals.
Although there is a docker image to containerize ROS processes, Gazebo should be installed on your host machine to avoid problems with using graphical resources. It is preferable to install the full version of ROS Noetic to avoid problems with gazebo_ros_pkgs dependencies.
To install both metapackages, just run
sudo apt install ros-noetic-desktop-full
Now, put the ROS Noetic setup.bash script in your bashrc file to load the ROS environment variables for each new terminal opening
echo "source /opt/ros/noetic/setup.bash" >> ~/.bashrc & source ~/.bashrc
Note: Use alias instead if you have multiple ROS workspaces or ROS1 and ROS2 versions on the same computer!
Open a terminal in your Host machine and run the ROS master process
roscore
Now, in a new terminal, execute the listener node of the roscpp_tutorial package
rosrun roscpp_tutorial listener
Open more one terminal and run the talker node on the docker container executing first the .run-container.sh
./run-container.sh
and thus the node
rosrun roscpp_tutorial talker
If everything is ok, you will see ROS INFO messages like the section above.
To see Gazebo running, run
roslaunch gazebo_ros empty_world.launch
and an empty Gazebo world has to appear
The ORB-SLAM3 method performs visual SLAM with three different sources of sensors: Monocular Camera, Depth Cameras (RGBD), and Monocular Camera + IMU. Therefore, this package has three launch files.
The monocular and RGBD launch files have the common arguments enable_pangolim, enable_rviz, and enable_trajectory_servers that are false by default. Pangolin is the ROS-independent method of visualization provided by ORB-SLAM3 core, while RViz is the standard visualization tool in the ROS community. enable_trajectory_servers enables the package [hector_trajectory_server] to generate ROS Paths.
The launch file for monocular + IMU has no arguments and no visualization tool enabled. Because the computational process of this approach is crashing, there wasn't enough time to configure the launch file for this approach.
Therefore, to launch ORB-SLAM3 nodes, execute the following commands.
- ORB-SLAM3 Monocular:
roslaunch vslam_comparison orb_slam3_mono.launch enable_rviz:=true enable_pangolim:=true enable_trajectory_servers:=true
- ORM-SLAM3 RGBD:
roslaunch vslam_comparison orb_slam3_rgbd.launch enable_rviz:=true enable_pangolim:=true enable_trajectory_servers:=true
- ORB-SLAM3 Monocular + IMU:
roslaunch vslam_comparison orb_slam3_mono_inertial.launch
TODO