Contents

  1. Description
  2. Demo

DEPRECATED: Use http://www.pointclouds.org/documentation/tutorials/vfh_recognition.php instead.

Description

This package implements a variant of the global VFH (Viewpoint Feature Histogram) descriptor, as presented in R.B. Rusu, G. Bradski, R. Thibaux, J. Hsu. Fast 3D Recognition and Pose Using the Viewpoint Feature Histogram, Proceedings of the 23rd IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Taipei, Taiwan, October 18-22, 2010.

The computational steps usually are along the lines of:

  • point cloud cluster extraction (e.g., through segmentation)
  • estimation of surface normals at each point
  • estimation of VFH global descriptors (one per cluster)
  • build a kd-tree of VFH signatures in some metric space (e.g., Chi-Square)
  • use the kd-tree for nearest neighbor searches to identify new (unseen) clusters.

Demo

Assuming that the package is built, the demos subdirectory contains a number of scripts that illustrate how to use VFH signatures for point cloud cluster classification. A number of example datasets are already bundled with the package and should be available in demos/data.

The following explains what every script in demos does:

  • 0.estimate_vfh_and_normals.sh needs estimate_normals_and_vfh.launch launched first, e.g.:

$ roslaunch vfh_cluster_classifier estimate_normals_and_vfh.launch

and for a given directory containing point cloud clusters, it will estimate the surface normals and VFH global descriptors;

  • 0.estimate_vfh.sh needs estimate_vfh.launch launched first, e.g.:

$ roslaunch vfh_cluster_classifier estimate_vfh.launch

and for a given directory containing point cloud clusters with surface normals at each point, it will estimate the VFH global descriptors;

  • 1.build_tree.sh uses the estimated VFH global descriptors to build a kd-tree representation in some given metric space;

  • 2.nearest_neighbors.sh uses the previously created kd-tree and displays the nearest neighbors in the chosen VFH metric space for one cluster per directory (from demos/data);

  • 3.view_histograms.sh draws the VFH signatures for one cluster per directory (from demos/data).

For a quick test, just run:

$ bash 1.build_tree.sh
$ bash 2.nearest_neighbors.sh

in demos. You should see something along the lines of:

ex1.jpg ex2.jpg

Here, the inlier distance threshold was set to 100. All shapes with distances larger than that are considered outliers.

Wiki: vfh_cluster_classifier (last edited 2011-05-19 17:24:22 by RaduBogdanRusu)