Limo

CALL FOR HACKATHON: Against all my expectations my work still invokes the interest of many. Since it aged quite a bit now, I would love to organize a Hackathon were people who love to code such as me take part to refurbish this repo. Possible things to do are:

1. Make a standalone api that works without ros
2. Make a python api (already started)
3. Bring repo to C++20 as much as possible (basis is ros indepedant app)
4. constexpr everything (as much as possible)
5. ROS2
6. Remake of the lidar point depth extraction module

But for this I need YOUR HELP! Your reward will be a lot of fun, working together on a project with experienced devs and of course a contribution record (that looks pretty neat in applications ;) ). If you are interested, please contact me per mail or write an issue. If there is more than 2 people plus me, I will organize a date :)

Lidar-Monocular Visual Odometry. This library is designed to be an open platform for visual odometry algortihm development. We focus explicitely on the simple integration of the following key methodologies:

• Keyframe selection
• Landmark selection
• Prior estimation
• Depth integration from different sensors.
• Scale integration by groundplane constraint.

The core library keyframe_bundle_adjustment is a backend that should faciliate to swap these modules and easily develop those algorithms.

• It is supposed to be an add-on module to do temporal inference of the optimization graph in order to smooth the result

• In order to do that online a windowed approach is used

• Keyframes are instances in time which are used for the bundle adjustment, one keyframe may have several cameras (and therefore images) associated with it

• The selection of Keyframes tries to reduce the amount of redundant information while extending the time span covered by the optimization window to reduce drift

• Methodologies for Keyframe selection:

  • Difference in time
  • Difference in motion

• We use this library for combining Lidar with monocular vision.

• Limo2 on KITTI is LIDAR with monocular Visual Odometry, supported with groundplane constraint

• Video: https://youtu.be/wRemjJBjp64

• Now we switched from kinetic to melodic

Details

This work was accepted on IROS 2018. See https://arxiv.org/pdf/1807.07524.pdf .

If you refer to this work please cite:

@inproceedings{graeter2018limo,
  title={LIMO: Lidar-Monocular Visual Odometry},
  author={Graeter, Johannes and Wilczynski, Alexander and Lauer, Martin},
  booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={7872--7879},
  year={2018},
  organization={IEEE}
}

Please note that Limo2 differs from the publication. We enhanced the speed a little and added additional groundplane reconstruction for pure monocular visual odometry and a combination of scale from LIDAR and the groundplane (best performing on KITTI). For information on Limo2, please see my dissertation https://books.google.de/books?hl=en&lr=&id=cZW8DwAAQBAJ&oi .

Installation

Docker

To facilitate the development I created a standalone dockerfile.

• Install docker https://phoenixnap.com/kb/how-to-install-docker-on-ubuntu-18-04
• build docker:

# This is where you put the rosbags this will be available at /limo_data in the container
mkdir $HOME/limo_data
cd limo/docker
docker-compose build limo

• You can run the docker and go to the entrypoint with

docker-compose run limo bash

Go to step Run in this tutorial and use tmux for terminals.

• You can invoke a jupyter notebook with a python interface for limo with

docker-compose up limo

and open the suggested link from the run output in a browser.

Semantic segmentation

The monocular variant expects semantic segmentation of the images. You can produce this for example with my fork from NVIDIA's semantic segmentation:

1. Clone my fork

git clone https://github.com/johannes-graeter/semantic-segmentation

2. Download best_kitti.pth as described in the README.md from NVIDIA and put it in the semantic-segmentation folder

3. I installed via their docker, for which you must be logged in on (and register if necessary) https://ngc.nvidia.com/

4. Build the container with

docker-compose build semantic-segmentation

5. Run the segmentation with

docker-copmose run semantic-segmentation

Note that without a GPU this will take some time. With the Nvidia Quadro P2000 on my laptop i took around 6 seconds per image.

Requirements

In any case:

• ceres:
  • follow the instructions on http://ceres-solver.org/installation.html
  • you will need sudo make install to install the headers.
  • tested with libsuitesparse-dev from standard repos.
• png++:

 sudo apt-get install libpng++-dev

• install ros:
  • follow the instructions on https://wiki.ros.org/melodic/Installation.
  • you will need to install ros-perception (for pcl).
  • don't forget to source your ~/.bashrc afterwards.
• install catkin_tools:

sudo apt-get install python-catkin-tools

• install opencv_apps:

sudo apt-get install ros-melodic-opencv-apps

• install git:

sudo apt-get install git

Build

• initiate a catkin workspace:

  cd ${your_catkin_workspace}
  catkin init

• clone limo into src of workspace:

  mkdir ${your_catkin_workspace}/src
  cd ${your_catkin_workspace}/src
  git clone https://github.com/johannes-graeter/limo.git

• clone dependencies and build repos

  cd ${your_catkin_workspace}/src/limo
  bash install_repos.sh

• unittests:

  cd ${your_catkin_workspace}/src/limo
  catkin run_tests --profile limo_release

Run

• get test data Sequence 04 or Sequence 01. This is a bag file generated from Kitti sequence 04 with added semantic labels.

• in different terminals (for example with tmux)

  1. roscore
  2. rosbag play 04.bag -r 0.1 --pause --clock

  3. source ${your_catkin_workspace}/devel_limo_release/setup.sh
     roslaunch demo_keyframe_bundle_adjustment_meta kitti_standalone.launch

  4. unpause rosbag (hit space in terminal)
  5. rviz -d ${your_catkin_workspace}/src/demo_keyframe_bundle_adjustment_meta/res/default.rviz

• watch limo trace the trajectory in rviz :)

• Before submitting an issue, please have a look at the section Known issues.

Known issues

• Unittest of LandmarkSelector.voxel fails with libpcl version 1.7.2 or smaller (just 4 landmarks are selected). Since this works with pcl 1.8.1 which is standard for ros melodic, this is ignored. This should lower the performance of the software only by a very small amount.