Paper: [arxiv] / Videos: [video1], [video2], [video3], [video4], [video5], [video6]
Authors: Maxime Ferrera, Alexandre Eudes, Julien Moras, Martial Sanfourche, Guy Le Besnerais (maxime.ferrera@gmail.com / first.last@onera.fr).
OV²SLAM is a fully real-time Visual SLAM algorithm for Stereo and Monocular cameras. A complete SLAM pipeline is implemented with a carefully designed multi-threaded architecture allowing to perform Tracking, Mapping, Bundle Adjustment and Loop Closing in real-time. The Tracking is based on an undirect Lucas-Kanade optical-flow formulation and provides camera poses estimations at the camera's frame-rate. The Mapping works at the keyframes' rate and ensures continuous localization by populating the sparse 3D map and minimize drift through a local map tracking step. Bundle Adjustment is applied with an anchored inverse depth formulation, reducing the parametrization of 3D map points to 1 parameter instead of 3. Loop Closing is performed through an Online Bag of Words method thanks to iBoW-LCD. In opposition to classical offline BoW methods, no pre-trained vocabulary tree is required. Instead, the vocabulary tree is computed online from the descriptors extracted in the incoming video stream, making it always suited to the currently explored environment.
If you use OV²SLAM in your work, please cite it as:
@article{fer2021ov2slam,
title={{OV$^{2}$SLAM} : A Fully Online and Versatile Visual {SLAM} for Real-Time Applications},
author={Ferrera, Maxime and Eudes, Alexandre and Moras, Julien and Sanfourche, Martial and {Le Besnerais}, Guy.},
journal={IEEE Robotics and Automation Letters},
year={2021}
}
OV²SLAM is released under the GPLv3 license. For a closed-source version of OV²SLAM for commercial purposes, please contact ONERA (https://www.onera.fr/en/contact-us) or the authors.
Copyright (C) 2020 ONERA
The library has been tested with Ubuntu 16.04 and 18.04, ROS Kinetic and Melodic and OpenCV 3. It should also work with ROS Noetic and OpenCV 4 but this configuration has not been fully tested.
OV²SLAM makes use of C++11 features and should thus be compiled with a C++11 or higher flag.
ROS is used for reading the video images through bag files and for visualization purpose in Rviz. ROS Installation
Make sure that the pcl_ros package is installed :
sudo apt install ros-distro-pcl-ros
or even
rosdep install ov2slam
Eigen3 is used throughout OV²SLAM. It should work with version >= 3.3.0, lower versions have not been tested.
OpenCV 3 has been used for the development of OV²SLAM, OpenCV 4 might be supported as well but it has not been tested. (Optional) The use of BRIEF descriptor requires that opencv_contrib was installed. If it is not the case, ORB will be used instead without scale and rotation invariance properties.
WATCH OUT By default the CMakeLists.txt file assumes that opencv_contrib is installed, set the OPENCV_CONTRIB setting to OFF in CMakeLists.txt if it is not the case.
A modified version of iBoW-LCD is included in the Thirdparty folder. It has been turned into a shared lib and is not a catkin package anymore. Same goes for OBIndex2, the required dependency for iBoW-LCD. Check the lcdetector.h and lcdetector.cc files to see the modifications w.r.t. to the original code.
Sophus is used for SE(3), SO(3) elements representation. For convenience, a copy of Sophus has been included in the Thirdparty folder.
Ceres is used for optimization related operations such as PnP, Bundle Adjustment or PoseGraph Optimization. For convenience, a copy of Ceres has been included in the Thirdparty folder. Note that Ceres dependencies are still required.
OpenGV can be used for Multi-View-Geometry (MVG) operations. The results reported in the paper were obtained using OpenGV.
For convenience, if OpenGV is not installed, MVG operations' alternatives are proposed with OpenCV functions.
Note that the performances might be lower in without OpenGV.
Clone the git repository in your catkin workspace:
cd ~/catkin_ws/src/
git clone https://github.com/ov2slam/ov2slam.git
For convenience we provide a script to build the Thirdparty libs:
cd ~/catkin_ws/src/ov2slam
chmod +x build_thirdparty.sh
./build_thirdparty.sh
WATCH OUT By default, the script builds obindex2, ibow-lcd, sophus and ceres. If you want to use your own version of Sophus or Ceres you can comment the related lines in the script. Yet, about Ceres, as OV²SLAM is by default compiled with the "-march=native" flag, the Ceres lib linked to OV²SLAM must be compiled with this flag as well, which is not the default case (at least since Ceres 2.0). The build_thirdparty.sh script ensures that Ceres builds with the "-march=native" flag.
If you are not interested in the Loop Closing feature of OV²SLAM, you can also comment the lines related to obindex2 and ibow-lcd.
(Optional) Install OpenGV:
cd your_path/
git clone https://github.com/laurentkneip/opengv
cd opengv
mkdir build
cd build/
cmake ..
sudo make -j4 install
Build OV²SLAM package with your favorite catkin tool:
cd ~/catkin_ws/src/ov2slam
catkin build --this
source ~/catkin_ws/devel/setup.bash
OR
cd ~/catkin_ws/
catkin_make --pkg ov2slam
source ~/catkin_ws/devel/setup.bash
Run OV²SLAM using:
rosrun ov2slam ov2slam_node parameter_file.yaml
Visualize OV²SLAM outputs in Rviz by loading the provided configuration file: ov2slam_visualization.rviz.
Both the Pinhole Rad-tan and Fisheye camera's models are supported. The models are OpenCV-based. If you use Kalibr for camera calibration, the equivalencies are:
- OpenCV "Pinhole" -> Kalibr "Pinhole Radtan"
- OpenCV "Fisheye" -> Kalibr "Pinhole Equidistant"
The stereo extrinsic parameters in the parameter files are expected to represent the transformation from the camera frame to the body frame (T_body_cam \ X_body = T_body_cam * X_cam). Therefore, if T_body_camleft is set as the Identity transformation, for the right camera we have: T_body_camright = T_camleft_camright. In Kalibr, the inverse transformation is provided (i.e. T_cam_body). Yet, Kalibr also provide the extrinsic transformation of each camera w.r.t. to the previous one with the field T_cn_cnm1. This transformation can be directly used in OV²SLAM by setting T_body_camleft = T_cn_cnm1 and T_body_camright = I_4x4.
Three directories are proposed within the parameter_files folder: accurate, average and fast. They all store the parameter files to be used with KITTI, EuRoC and TartanAir.
The accurate folder provides the parameters as used in the paper for the full method (i.e. OV²SLAM w. LC).
The fast folder provides the parameters as used in the paper for the Fast version of OV²SLAM.
The average folder is provided for convenience as an in-between mode.
TODO: add description of the parameters file
If you experience issues when running OV²SLAM (segfault exceptions, ...), it might be related to the "-march=native" flag. By default, OpenGV and OV²SLAM come with this flag enabled but Ceres does not. Making sure that all of them are built with or without this flag might solve your problem.