FlexMap Fusion

Overview

Offline Mapping Pipeline developed at the Institute of Automotive Technology

• ROS2 package for the improvement/extension of lanelet2 maps with data from OpenStreetMap and their georeferencing with GNSS data
• The package is designed for the application in an offline mapping pipeline developed at the Institute of Automotive Technology of the Technical University of Munich
• The package consists of the three modules "Map Alignment", "Map Conflation" and "Georeferencing", that are explained in more detail in the image below and the documentation of the single modules Overview of the single modules of the FlexMap Fusion tool
• The following functionalities are included
  • georeferencing of the lanelet2 map based on the transformation calculated by the vehicle ego trajectory from GNSS and the SLAM-trajectory
  • fusion of semantic information from OpenStreetMap into the lanelet2 map (generated from point cloud map data)
  • modular expandability to include additional semantic information
  • possibility to apply georeferencing step on the point cloud map resulting from the SLAM process and used for the generation of the lanelet map
  • extensive visualization within RVIZ2

🐋 Docker Setup

Package Design

This package is designed as a standalone ROS2 package. It was developed with ROS2 humble. For easier handling of dependencies, a docker environment is provided that sets up everything and builds the package. As the package is designed for use in combination with Autoware, the source code can also be build within the Autoware docker environment.

Setup

1. Clone the repository by running

   git clone [email protected]:TUMFTM/FlexMap_Fusion.git

2. Go to the root directory of the repository

   cd FlexMap_Fusion/

3. Build the docker image

   ./docker/build_docker.sh  

4. Run the container and mount your data by appending the directory containing your data:

   ./docker/run_docker.sh /your/local/directory/data

🖥 How to Use the Package

• the package contains two executables with corresponding ROS2 launch file:

  • lanelet2_osm
    • provide all functionality described in the publication and the pipeline overview
    • additional possibility to georeference the point cloud map corresponding to the lanelet map, but without its visualization
  • kiss_icp_georef
    • provides the possibility to georeference the SLAM poses and the corresponding point cloud map withouth the need of a lanelet map as input (-> no conflation with OpenStreetMap)
    • provides visualization of the point cloud map in RVIZ2

• in the following, the sections are split between the two executables (however, keep in mind that kiss_icp_georef just provides a subset of the functions of lanelet2_osm)

📄 lanelet2_osm

1. Necessary input parameters:

   • traj_path => path to GPS trajectory of the vehicle (format: txt-file with lat, lon)
   • poses_path => path to SLAM trajectory of the vehicle (KITTI-format, trajectories don't have to be synchronized over time)
   • map_path => path to lanelet2 map corresponding to trajectories (map can have missing elements/attributes, only when using node lanelet2_osm)
   • out_path => path to save the modified lanelet map (DEFAULT: /lanelet2_map.osm, only when using node lanelet2_osm)
   • if you want to georeference the point cloud map corresponding to the lanelet2 map with the same set of control points:
     • set the parameter transform_pcd in the config file to true
     • adjust the path to the point cloud map (parameter pcd_path)
     • the georeferenced point cloud map will be saved in the current working directory (if you'd like to specify a different path, see kiss_icp_georef

2. Start the package

   • it is recommended to directly use the provided ROS launch file as it starts the package itself and the visualization in RVIZ:
   • replace the filepaths and run the following command inside the docker container:

       ros2 launch flexmap_fusion lanelet2_osm.launch.py traj_path:=<path-to-GPS-trajectory> poses_path:=<path-to-SLAM-trajectory>  map_path:=<path-to-lanelet-map> out_path:=<path-to-save-output-map>

   • the launch file directly links to the corresponding parameter file in /config/.
   • To directly run the package with the provided test files from the Docker root directory, use the following command:

        ros2 launch flexmap_fusion lanelet2_osm.launch.py traj_path:=./src/flexmap_fusion/test/route_1_GPS.txt poses_path:=./src/flexmap_fusion/test/route1_pose_kitti.txt  map_path:=./src/flexmap_fusion/test/lanelet2_route_1.osm out_path:=lanelet2_map_georef.osm

3. Select control points

   • after the trajectories are loaded and the target trajectory is aligned to the master trajectory by the Umeyama algorithm, you are asked in the command window to select control points for the rubber-sheet transformation (the amount of points can be configured in the config file).
   • select the desired points using the Publish Point button in RVIZ and follow the instructions in the command window.

4. Inspect results

   • results of the rubber-sheet transformation & lanelet map are visualized
   • see table for explanation of single topics

Topic	Description
/lof/map/osm_map_markers	Street network downloaded from OpenStreetMap.
/lof/map/ll_map_markers	Lanelet2-map enriched with attributes from OpenStreetMap; Lanelets are colorized based on agreement between adjacent lanelets and lanes-tag of OpenStreetMap.
/lof/map/ll_map_new_markers	Lanelet2-map enriched with attributes from OpenStreetMap; Lanelets that are likely to be wrong were removed by the module "Deletion of Lanelet Fragments".
/lof/traj/traj_master_markers	Master trajectory -> to be defined in config-file (usually GNSS-trajectory)
/lof/traj/traj_target_markers	original target rajectory (usually SLAM-trajectory)
/lof/traj/traj_align_markers	target trajectory aligned to master with Umeyama-algorithm
/lof/traj/traj_rs_markers	target trajectory after rubber-sheet-transformation based on control points
/lof/rs/geom_markers	geometric information from rubber-sheeting (control points and constructed triangles)
/clicked_point	last 2 selected points by user to indicate chosen control points
/lof/confl/geom_markers	geometric information regarding conflation process (collapsed lanelet-map, buffers, matches)

• Inspect results and modify parameters if desired.

5. Manually finalize lanelet map
   • open a manual editor for lanelet2 maps (e.g. VectorMapBuilder) in parallel to RVIZ
   • import the exported map from out_path
   • close gaps in lanelet map and correct other mistakes based on visualization of map agreement with OpenStreetMap in RVIZ

📄 kiss_icp_georef

1. Necessary input parameters:

   • traj_path => path to GPS trajectory of the vehicle (format: txt-file with lat, lon)
   • poses_path => path to SLAM trajectory of the vehicle (KITTI-format, trajectories don't have to be synchronized over time)
   • pcd_path => path to pcd map corresponding to poses trajectory
   • pcd_out_path => path to saved the georeferenced point cloud map (DEFAULT: /pcd_map_georef.pcd)

2. Start the package

   • it is recommended to directly use the provided ROS launch file that starts the package itself and the visualization in RVIZ:

       ros2 launch flexmap_fusion kiss_icp_georef.launch.py traj_path:=<path-to-GPS-trajectory> poses_path:=<path-to-SLAM-trajectory>  pcd_path:=<path-to-pcd-map> pcd_out_path:=<path-to-save-pcd-map>

   • the launch file directly links to the corresponding parameter file in /config/.

3. Select control points

   • after the trajectories are loaded and the target trajectory is roughly aligned to the master trajectory you are asked in the command window to select control points for the rubber-sheet transformation (the amount of points can be configured in the config file).
   • select the desired points using the Publish Point button in RVIZ and follow the instructions in the console.

4. Inspect results

   • results of the rubber-sheet transformation & the resulting, transformed point cloud map are visualized.
   • see table for explanation of single topics

Topic	Description
/lof/traj/traj_master_markers	Master trajectory -> to be defined in config-file (either GNSS- or SLAM trajectory)
/lof/traj/traj_target_markers	original target rajectory -> depending on selected master trajectory (either GNSS- or SLAM trajectory)
/lof/traj/traj_align_markers	target trajectory aligned to master with Umeyama transformation or PCL ICP
/lof/traj/traj_rs_markers	target trajectory after rubber-sheet-transformation
/lof/rs/geom_markers	geometric information from rubber-sheeting (control points and constructed triangles)
/clicked_point	last 2 selected points by user to indicate chosen control point
/lof/rs/pcd_map	transformed point cloud map (only when using kiss_icp_georef)

• Inspect results and modify parameters if desired.

• the parameter files for both executables are located in /config
• see the comments within the single .param.yaml-files for detailed explanations on the parameters
• to modify the parameters inside the container and view their current value, python-executables are provided:
  • to view the current value of a parameter, run

     get_param.py <config_file_name> <param_name>

  inside the container.
  • to modify a parameter, run

     config_param.py <config_file_name> <param_name> <param_value> <param_type>

  inside the container. Supported parameter types are str, int, float and bool. See the current value for the right choice, otherwise the node will crash.
  • Example: Set amount of control points for rubber-sheeting:

     config_param.py lanelet2_osm.param.yaml rs_num_controlPoints 10 int

  Check effect:

     get_param.py lanelet2_osm.param.yaml rs_num_controlPoints

📈 Test Data

(GPS trajectory). The SLAM poses were generated by [KISS-ICP](https://github.com/PRBonn/kiss-icp) in combination with [interactive SLAM](https://github.com/SMRT-AIST/interactive_slam). The lanelet2-map was created manually with [VectorMapBuilder](https://tools.tier4.jp/feature/vector_map_builder_ll2/) by TieriV.

🔧 Modules

Detailed documentation of the functionality behind the single modules can be found below.

1. Geometric Alignment

2. Preprocessing

3. Matching

4. Conflation

5. Georeferencing

6. Analysis

📇 Contact Info

Maximilian Leitenstern, Institute of Automotive Technology, School of Engineering and Design, Technical University of Munich, 85748 Garching, Germany

Florian Sauerbeck, Institute of Automotive Technology, School of Engineering and Design, Technical University of Munich, 85748 Garching, Germany

Dominik Kulmer, Institute of Automotive Technology, School of Engineering and Design, Technical University of Munich, 85748 Garching, Germany

📃 Citation

If you use this repository for any academic work, please cite our original paper:

@misc{leitenstern2024flexmap,
      title={FlexMap Fusion: Georeferencing and Automated Conflation of HD Maps with OpenStreetMap}, 
      author={Maximilian Leitenstern and Florian Sauerbeck and Dominik Kulmer and Johannes Betz},
      year={2024},
      eprint={2404.10879},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}