Dense Visual SLAM for RGB-D Cameras (dvo_slam)

This repository contains a DVO-SLAM fork (ROS Kinetic) with modifications for our BMVC 2017 paper "Efficient Online Surface Correction for Real-time Large-Scale 3D Reconstruction". In particular, this code also outputs the current keyframe graph to disk when the keyframe map changes.

Packages

• dvo_core Core implementation of the motion estimation algorithm.
• dvo_ros Integration of dvo_core with ROS.
• dvo_slam Pose graph SLAM system based on dvo_core and integration with ROS.
• dvo_benchmark Integration of dvo_slam with TUM RGB-D benchmark, see http://vision.in.tum.de/data/datasets/rgbd-dataset.

Installation

Setup ROS environment and create a ROS catkin workspace:

# source ROS environment
source /opt/ros/kinetic/setup.bash

# create ROS workspace folder
mkdir ros_catkin_ws
cd ros_catkin_ws

# create catkin workspace and source setup .sh file
catkin_make
source devel/setup.bash

Checkout and build the DVO-SLAM source code:

# clone repository
cd src/
git clone https://github.com/robmaier/dvo_slam_correctablefusion.git

# create TUM benchmark output directories
cd dvo_slam_correctablefusion/dvo_benchmark
mkdir dvo_slam_correctablefusion/dvo_benchmark/output/
mkdir dvo_slam_correctablefusion/dvo_benchmark/output/graph/
cd ..

# build package (and workspace) using catkin_make
catkin_make

Dataset

Download one of the TUM RGB-D Benchmark sequences:

# go into data folder
cd src/dvo_slam_correctablefusion/data/
# download dataset
wget https://vision.in.tum.de/rgbd/dataset/freiburg3/rgbd_dataset_freiburg3_long_office_household.tgz
# extract data
tar -xvzf rgbd_dataset_freiburg3_long_office_household.tgz
mv rgbd_dataset_freiburg3_long_office_household fr3_office
# associate color and depth images using their timestamps
python associate.py fr3_office/rgb.txt fr3_office/depth.txt > fr3_office/assoc.txt
# go back to parent folder
cd ../../../

Usage

Run roscore (in a new different terminal):

source /opt/ros/kinetic/setup.bash
roscore

Estimate camera trajectory for the downloaded RGB-D dataset:

roslaunch dvo_benchmark benchmark.launch dataset:=$PWD/src/dvo_slam_correctablefusion/data/fr3_office

We calculate the camera tracking accuracy and plot the differences between the groundtruth trajectory and the estimated trajectory as follows:

cd src/dvo_slam_correctablefusion/data/fr3_office
python ../evaluate_ate.py groundtruth.txt ../../dvo_benchmark/output/trajectory.txt --plot plot.png --verbose

You can also use RVIZ for real-time visualization:

source /opt/ros/kinetic/setup.bash

# start RVIZ
rosrun rviz rviz

In the GUI,

• Start RVIZ
• Set the Fixed Frame (Global Options) to /world
• Add an Interactive Marker display and set its Update Topic to /dvo_vis/update
• Optional: add a PointCloud2 display and set its Topic to /dvo_vis/cloud

Publications

The following publications describe the approach:

• Dense Visual SLAM for RGB-D Cameras (C. Kerl, J. Sturm, D. Cremers), In Proc. of the Int. Conf. on Intelligent Robot Systems (IROS), 2013.
• Robust Odometry Estimation for RGB-D Cameras (C. Kerl, J. Sturm, D. Cremers), In Proc. of the IEEE Int. Conf. on Robotics and Automation (ICRA), 2013
• Real-Time Visual Odometry from Dense RGB-D Images (F. Steinbruecker, J. Sturm, D. Cremers), In Workshop on Live Dense Reconstruction with Moving Cameras at the Intl. Conf. on Computer Vision (ICCV), 2011.

License

The packages dvo_core, dvo_ros, dvo_slam, and dvo_benchmark are licensed under the GNU General Public License Version 3 (GPLv3), see http://www.gnu.org/licenses/gpl.html.