Fast Direct Stereo Visual SLAM

Related Publications

Direct Sparse Odometry, J. Engel, V. Koltun, D. Cremers, In IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2018
Extending Monocular Visual Odometry to Stereo Camera System by Scale Optimization, J. Mo and J. Sattar, In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019
A Fast and Robust Place Recognition Approach for Stereo Visual Odometry Using LiDAR Descriptors, J. Mo and J. Sattar, In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020

Note

This implementation is tested in Ubuntu 20.04 with ROS Noetic.

Dependencies

ROS, PCL, g2o, and DSO (OpenCV, Pangolin)

For g2o, DSO, and Pangolin, we strongly recommend the (old) versions in the provided dependencies.zip for smooth install and reasonable results.

Install Pangolin

sudo apt install libgl1-mesa-dev libglew-dev cmake 
sudo apt install libegl1-mesa-dev libwayland-dev libxkbcommon-dev wayland-protocols
cd Pangolin
mkdir build && cd build
cmake ..
make -j4
sudo make install # or set LD_LIBRARY_PATH locally

Install DSO

sudo apt install libsuitesparse-dev libeigen3-dev libboost-all-dev libopencv-dev
cd dso
mkdir build && cd build
cmake ..
make -j4

Install g2o

cd g2o
mkdir build && cd build
cmake -DBUILD_WITH_MARCH_NATIVE=ON .. # use the flag to avoid possible double-free error
make -j4
sudo make install # or set LD_LIBRARY_PATH locally

Install

Link to the external DSO library:

export DSO_PATH=[PATH_TO_DSO] (e.g., export DSO_PATH=~/Workspace/dso)

or set the DSO_PATH in CMakeLists.txt.

Install direct_stereo_slam using ros/catkin

cd catkin_ws/src
git clone https://github.com/IRVLab/direct_stereo_slam.git
cd ..
catkin_make

Usage

Preparation

Calibrate stereo cameras with format of cams. T_stereo is the pose of camera0 in camera1 coordinate, rememeber to put a small number in T_stereo[2,2] for numerical stability if images are stereo pre-calibrated. Refer to DSO for more details of intrisic parameters.
Create a launch file with the format of sample.launch.
For the evaluation on KITTI and Malaga dataset, we follow kitti2bag to generate the bag files.

Parameters (in launch file)

scale_opt_thres: scale optimization accept threshold (e.g., 15.0)
lidar_range: imitated LiDAR scan range, set to -1 to disable loop closure (e.g., 40.0 meters)
scan_context_thres: Scan Context threshold for a potential loop closure (e.g., 0.33)

Run

roslaunch direct_stereo_slam [YOUR_LAUNCH_FILE]

Ctrl-C to terminate the program, the final trajectory (dslam.txt) will be written to ~/.ros folder.

Output file (in ~/.ros folder)

dslam.txt: final trajectory [incoming_id, x, y, z];
sodso.txt: the trajectory without loop closure, output for comparision.

IRVLab / direct_stereo_slam