3D regitration of point clouds grabbed via Kinect

3D printed suitcase

Dependencies

• pcl 1.7

Build

• git clone https://github.com/DanieleVeri/pcl_registration
• mkdir pcl_registration/build
• cd pcl_registration/build
• cmake ..
• make

Run

• ./capture <folder> launch the point cloud grabber and store .pcd in <folder>
• ./generate_model <folder> launch the model generator, with .pcd in <folder> as input

Algorithm overview

Parameter tuning: hyperparameter.json

1. Depth images acquired are preprocessed:

• z filtered
• downsampeld
• main plain removal
• outlier removal
• keypoint extraction
• sift descriptor extraction FPFH

2. Pairwise registration

• many SampleConsensusInitialAlignment computed
• clustering of computed 4x4 transformation matrices
• main centroid used as init alignement
• ICP refinement
• main objects extracted from the pointclouds
• then aligned according the computed transformation

3. Cloud smoothing and mesh generation

• output smoothed point cloud
• output vtk mesh

TODO

• multithread registration
• backtrack axlong global alignement