Skip to content
release
Switch branches/tags
Code
This branch is up to date with release.
Contribute

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 

VI-MEAN

Visual-Inertia-fusion-based Monocular dEnse mAppiNg

This repository corresponds to our paper on ICRA 2017:

Real-time Monocular Dense Mapping on Aerial Robots using Visual-Inertial Fusion

Zhenfei Yang, Fei Gao, and Shaojie Shen

paper and video

How to view the 3D model

The following rosbags might be helpful to take a close look at the dense 3D model produced by our system.

https://www.dropbox.com/s/pwbw0cz1i26sjba/2016-09-18-15-46-27-part1.bag?dl=0

https://www.dropbox.com/s/5uboefps9js13bs/2016-09-18-16-07-25-part2.bag?dl=0

Topic of Meshed dense map: /Chisel/full_mesh

Topic of estimated path: /self_calibration_estimator/path

Note: Unfortunately, the model is too dense to be recorded real-time using our onboard computers. ROS randomly throw data with limited IO, resulting very low frequency of the messages in the bags.

We emphasize the map received by controller is updated at 10Hz. The frequency of map generation really matters a lot for autonomous systems.

How to compile

  1. Choose correct CUDA_NVCC_FLAGS in stereo_mapper/CMakeLists.txt
  2. Install the following dependencies:
  • OpenCV
  • Eigen
  • Ceres (http://ceres-solver.org/)
  • Modified OpenChisel (already included in this repo), which requires PCL compiled with c++11. Please follow OpenChisel's instruction.
  • Subset of camodocal (already inclued in this repo with name of camera_model) to support wide-angle lens.
  1. catkin_make

How to run it with our data

  1. Download sample.bag via: http://uav.ust.hk/storage/sample.bag
  2. roslaunch stereo_mapper sample_all.launch
  3. rosbag play path_to_the_bag/sample.bag
  4. Use rviz with your config file sample.rviz to visualize

How to run it with your data

  1. Calibrate your camera using camera_model to get the calibration file that looks like sample_camera_calib.yaml
  2. Modify the launch file to fit your topic names and the path to your calibration file
  3. Follow the above steps

Licence

VI-MEAN is licensed under the GNU General Public License Version 3 (GPLv3), see http://www.gnu.org/licenses/gpl.html.

For commercial purposes, a more efficient version under different licencing terms is under development.

About

Visual-Inertia-fusion-based Monocular dEnse mAppiNg

Resources

License

Releases

No releases published

Packages

No packages published