Code of a manuscript, submitted to IEEE Transactions on Intelligent Transportation Systems. video

If have some questions on this project, please contact Huan Yin [email protected]

@article{yin20193d,
  title={3D LiDAR Map Compression for Efficient Localization on Resource Constrained Vehicles},
  author={Yin, Huan and Wang, Yue and Tang, Li and Ding, Xiaqing and Huang, Shoudong and Xiong, Rong},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  status={Accept}
}

After compression by RF, selected points are colored in red on the origin map
     

This project is compiled on ros and Ubuntu. And use catkin_make in your workspace to build this project. Some dependencies are needed to be prepared:

• pcl
• libpointmatcher (an old version in 2018.09)
• Gurobi

• cfg
  contains configuration files: .rviz & filter.yaml, for visualization & configuration`

• gurobi
  integer linear programming code, written in Matlab

• launch
  roslaunch files to start-up ros nodes, C++

• prepare
  sparcify the dense poses, generate keep.txt

• random_forest
  train & test, learning part, copy from GitHub, thanks to ranger

• src
  .cpp files using ros api & libpointmatcher

• images
  some pictures of results long time ago

• prepare poses and scans
  Maltab files, get the sparser poses for robot stopping cases, distance threshold used for more uniform of poses
  input: origin poses or trajectories
  output: uniform poses or trajectories

• before
  previous try on programming, no use

• q_ILP_lamda
  implementation for weighted ILP on map compression, including spliting and merging iteratively

• build
  training and testing in bash files

• formatTransfer
  change the cloud files in a folder (.vtk to .ply for example)
  input:　the folder address
  output: new format in this folder

• mapCheck
  check the map in Rviz
  __input:__map and poses (trajectory)

• normalViewer
  show the generated surface normals
  __input:__the map and some params

• scanRegister
  use origin raw scans and popses to get the origin full map, filter raw scans to reduce large size
  input:　uniform poses and origin scans
  output: dense origin full map

• mapFilter
  filter the dense origin full map using Octree Grid, and generate surface normals
  input: dense origin full map
  output: origin full map

• mapScoring
  score each laser point in map according to the strategy in paper, which is also for observation filter as a comparison input: origin full map, laser scans and poses
  output: scored origin full map

• genWeightVector
  get the weight vector for integer linear programming
  input: scored origin full map
  output: all points' weight in a saved txt

• genVisMatrix
  get the Visbility Matrix for integer linear programming
  input: origin full map, laser scans and poses (same as mapScoring)
  output: saved vis-files one by one in a folder

• mapCutterTraj
  split/Cut the whole Map into two parts: train or test, according to the nearest search on trajectory
  input:　oringn full map, poses and cutPoition
  output: train map, test map and labeled full map

• generateAllFeatures
  generate point feature for training or testing
  input:　oringn full map, poses
  output: map with features in descriptors and features in a saved txt file

• loadFinalResults
  annotate the origin full map with the final learned result
  input: origin full map and learned result in txt file
  output: labeled learned result in cloud file

• loadProResult
  annotate the origin full map with the supervised result by programming
  input:　origin file map and programmed results in files in a folder
  output: labeled programmed result in cloud file

• mapCompresser
  generate compressed map according to the annotated map on descriptor
  input:　annotated map by learning or programming, and the descriptor name
  output: compressed map

• compareDistribution
  get the nearest distances between two compressed clouds
  input:　two maps
  output: distances in a saved txt file

• errorDistribution
  get the nearest distances between salient & salient_predicted in cloud (learning & programming)
  input:　 one map with two descriptors (s & s_predicted)
  output: distances in a saved txt file

• locTest
  localization test by registration
  input:　 point cloud map, laser scans, trajectory(for position initial)
  output: pose saved txt, time saved txt, icp iteration saved txt

• clusterMap
  achieve map compression by clustering point clouds
  input:　 map, and some params
  output: saved clustered map

• voxelMap
  achieve map compression by voxel grid (similar as octree grid)
  input:　 map
  output: saved sparse map

• randomSampleMap
  achieve map compression by random selection
  input:　 map
  output: saved random map

• selectByObserver
  achieve map compression by thresholding the observation count
  input:　 map
  output: saved threshold map

• splitCloudForSaliency
  split map to sub-maps for geometry saliency detection
  input:　 map, and the trajectory
  output: sub-maps in a folder

• saliencyMap
  genrate saliency map on point clouds
  input:　 the folder contains sub-maps,
  output: saliency evaluation saved in txt files ina folder

• mergeCloudForSaliency
  merge the saliency and sub-maps into a whole map
  input:　 map, trajectory and saliency files
  output: one global map with saliency evaluation

MIT