Open REALM: Real-time Aerial Localization and Mapping

This is the repository for Open REALM, a real-time aerial mapping framework. It is currently in alpha state, so don't expect a polished and bugfree application.

Feel free to fork and contribute. Let's make mapping fast again! :)

Demonstration

Introduction

The proposed framework stands on the shoulders of giants, namely the open source implementation for visual SLAM and stereo reconstruction. Please read the references below.

For a detailed description of the underlying ideas refer to my thesis:

https://drive.google.com/open?id=1zXxuCVN4wzlC-ZAZBz5qI4wLkv1b6CbH

Beforehand

The pipeline is designed for multirotor systems with downlooking 2-axis gimbal stabilized camera and alignment of the magnetic heading of the UAV with the image's negative y-axis. It should be easily extendable to other cases in the future, but at the moment I won't recommend to use it on any other.

What it can do

2D maps based on lat, lon, alt and heading informations only
2D maps based on lat, lon, alt and visual SLAM pose estimation
2.5D elevation maps based on lat, lon, alt and visual SLAM pose estimation using sparse cloud
2.5D elevation maps based on lat, lon, alt and visual SLAM pose estimation using GPU stereo reconstruction
All of the above from a single or multiple UAVs

What NOT to expect

Transportation of your data from the UAV to the ground based on any other than ROS
Don't expect magic, mapping with visual informations can only be done in regions with actual features. The final map will only be as accurate as the chosen implementation for pose estimation, densification and fusion of both in the mosaic.
Polished, ready-to-use code for commercial applications. It is research code that has developed over a long time and is therefore sometimes ugly and buggy as hell. But I am working on it :)

Dependencies

Linux (tested with Ubuntu 16.04)

ROS Kinetic with OpenCV 3.3.1

-> Refer to http://wiki.ros.org/kinetic/Installation

Pangolin for ORB SLAM 2

-> https://github.com/stevenlovegrove/Pangolin

QT5 for RVIZ plugin

-> Refer to http://doc.qt.io/qt-5/qt5-intro.html

Exiv2 installation from source for image reading and writing with meta infos

-> https://github.com/Exiv2/exiv2

# Catkin tools for build process
sudo apt-get install python-catkin-tools

# Eigen3 for several linear algebra problems
sudo apt-get install libeigen3-dev

# Gdal library for conversions between UTM and WGS84
sudo apt-get install gdal-bin

# Cgal library for delauney 2.5D triangulation and mesh creation
sudo apt-get install libcgal-dev
sudo apt-get install libcgal-qt5-dev

# PCL for writing point clouds and mesh data
sudo add-apt-repository ppa:v-launchpad-jochen-sprickerhof-de/pcl
sudo apt-get update
sudo apt-get install libpcl-all

Optional Dependencies

CUDA for stereo reconstruction with plane sweep lib

-> Refer to https://developer.nvidia.com/cuda-downloads?target_os=Linux

Installation

Linux (tested with Ubuntu 16.04)

# Create and init a catkin workspace
mkdir -p catkin_ws/src
cd catkin/src

# Clone Open REALM git and compile
git clone https://github.com/laxnpander/OpenREALM.git

Option 1: Configuration WITHOUT cuda

# Make sure you are in your catkin_ws, not src
# Configure catkin and cmake, blacklist cuda dependent packages
catkin init --workspace .
catkin config --blacklist psl --cmake-args -DCMAKE_BUILD_TYPE=Release
catkin build

Option 2: Configuration WITH cuda

# Make sure you are in your catkin_ws, not src
# Configure catkin and cmake, no blacklisted packages, densifier with cuda
catkin config --blacklist "" --cmake-args -DCMAKE_BUILD_TYPE=Release -DDENSIFIER_WITH_CUDA=True
catkin build

Quick Start

Step 1: Download the test dataset:

https://drive.google.com/open?id=1-2h0tasI4wzxZKLBbOz3XbJ7f5xlxlMe

Step 2: Unzip the dataset with a tool of your choice, e.g.

tar -xvzf open_realm_edm_dataset.tar.gz

Step 3: We provided as well a set of configuration files in realm_ros/profiles as the corresponding launch files in realm_ros/launch to run the test dataset. The only thing you have to do is modify the path in the launch file:

node pkg="realm_ros" type="realm_exiv2_grabber" name="realm_exiv2_grabber" output="screen"
    param name="config/id" type="string" value="$(arg camera_id)"/>
    param name="config/input" type="string" value="PUT THE TEST DATASET'S ABSOLUTE PATH HERE"/>
    param name="config/rate" type="double" value="10.0"/>
    param name="config/profile" type="string" value="alexa"/>
/node

Note: The exiv2 grabber node reads images and exiv2 tags from the provided folder and publishes them for the mapping pipeline.

Step 4: Launch the pipeline you want to use.

GNSS only mapping:

roslaunch realm_ros alexa_gnss.launch

2D mapping with visual SLAM:

roslaunch realm_ros alexa_noreco.launch

2.5D mapping with visual SLAM and surface reconstruction:

roslaunch realm_ros alexa_reco.launch

References

Visual SLAM

[1] Raúl Mur-Artal, J. M. M. Montiel and Juan D. Tardós. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147-1163, 2015. (2015 IEEE Transactions on Robotics Best Paper Award).

Stereo Reconstruction

[2] Christian Häne, Lionel Heng, Gim Hee Lee, Alexey Sizov, Marc Pollefeys, Real-Time Direct Dense Matching on Fisheye Images Using Plane-Sweeping Stereo, Proc Int. Conf. on 3D Vison (3DV) 2014

Other

[3] P. Fankhauser and M. Hutter, "A Universal Grid Map Library: Implementation and Use Case for Rough Terrain Navigation", in Robot Operating System (ROS) – The Complete Reference (Volume 1), A. Koubaa (Ed.), Springer, 2016.

[4] T. Hinzmann, J. L. Schönberger, M. Pollefeys, and R. Siegwart, "Mapping on the Fly: Real-time 3D Dense Reconstruction, Digital Surface Map and Incremental Orthomosaic Generation for Unmanned Aerial Vehicles"

konanrobot / OpenREALM