ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | ros @ Robotics Stack Exchange
Ask Your Question

Plot a Gaussian 3D representation with markers in rviz

asked 2011-08-31 21:52:25 -0500

updated 2011-11-15 23:59:02 -0500

Hi people! I'm thinking of plotting a 3D Gaussian Model with a marker defined by a mean vector and a covariance matrix. Is there any rviz plugin to render this? If not, is it possible to do so plotting a marker?

I think it could be possible just plotting a sphere centered in the mean vector and apply to it a linear deformation using the covariance matrix.

Any idea or suggestion about how to address this problem?

EDITED: after the responses I made a blog entry to illustrate the problem:

edit retag flag offensive close merge delete

3 Answers

Sort by » oldest newest most voted

answered 2011-09-13 07:12:08 -0500

hersh gravatar image

Dornhege's link to the other discussion (ending ultimately with an enhancement request ticket for rviz) is certainly correct.

However to answer the question more specifically: yes you can do this with a Marker. Specify marker type SPHERE, set the marker.scale.x, y, and z values to the length you want each axis of the "sphere" to be. If the scale x, y, and z values are not all equal it will look like an ellipsoid instead of a sphere. Set marker.pose.orientation to give it an orientation and marker.pose.position to the location of the center, and you are all set.

I'm not the one to ask about the math to go from a covariance matrix to a rotated ellipsoid, but I know you can show rotated ellipsoids in rviz.

edit flag offensive delete link more


You just need eigenvectors and eigenvalues from the covariance matrix.
dornhege gravatar image dornhege  ( 2011-09-13 09:46:47 -0500 )edit

answered 2011-09-15 03:35:19 -0500

updated 2011-09-20 03:00:14 -0500

The eigenvectors represent the rotation matrix of the ellipsoid while the eigenvalues represent the scale of the sphere through the rotation frame axes. Here you can see a blog post with a more detailed code and a video with the results.

Here you can see a chunk of python code about how to solve the problem (thanks to hersh and dornhege for the suggestions):

enter code here
(eigValues,eigVectors) = numpy.linalg.eig (covMat)


rot = PyKDL.Rotation (eigx_n,eigy_n,eigz_n)
quat = rot.GetQuaternion ()

#painting the Gaussian Ellipsoid Marker
marker.pose.orientation.x =quat[0]
marker.pose.orientation.y = quat[1]
marker.pose.orientation.z = quat[2]
marker.pose.orientation.w = quat[3]
marker.scale.x = eigValues[0]
marker.scale.y = eigValues[1]
marker.scale.z =eigValues[2]
edit flag offensive delete link more


Note: although the eigenvalues are guaranteed to be positive, the eigenvector matrix may either have a determinant of +1 or -1. Check the eigenvectors carefully before treating it as a rotation matrix. By the way, it's customary to use sqrt(eigvalues) for scale.

dllu gravatar image dllu  ( 2016-12-06 17:54:19 -0500 )edit

answered 2011-09-01 00:24:52 -0500

dornhege gravatar image

I think this question is relevant here (seem it is not directly supported currently).

edit flag offensive delete link more

Question Tools


Asked: 2011-08-31 21:52:25 -0500

Seen: 5,286 times

Last updated: Nov 15 '11