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

how to identify an object from the generated point cloud?

asked 2011-10-05 17:22:56 -0600

manns gravatar image

updated 2014-04-20 14:09:35 -0600

ngrennan gravatar image

point cloud and ros

edit retag flag offensive close merge delete

3 Answers

Sort by ยป oldest newest most voted

answered 2011-10-06 01:54:33 -0600

raphael favier gravatar image

updated 2011-10-06 01:54:52 -0600


I had a few shots with VFH features recognitiom which work only on 3d information. Check these pcl tutorials:

How 3D Features work in PCL

Estimating VFH signatures for a set of points

Cluster Recognition and 6DOF Pose Estimation using VFH descriptors

It kind of gives good results, provided that you took enough snapshots from enough viewpoints of your objects during the learning phase.

I noticed that someone pushed it a level farther, by learning objects directly from 3d models and clustering the observed point clouds.

This work is advertised in this video and hosted in the ROS packagevfh_recognition.

As other said, object recognition stays a hard problem.

I hope it helps


edit flag offensive delete link more

answered 2011-10-06 01:43:14 -0600

Constantin S gravatar image

updated 2011-10-06 01:45:47 -0600

Start with Bayesian Probability, then read about Bayesian Error Rate. From there read about identifying simple properties of an object (such as height). Set a threshold on height, now you have separation. This is the absolute most basic parametric based object recognition algorithm, but it shows you the concepts of probability and error rates, which is the basis of all (I dare say) all object recognition methods. (parametric and non)

Source: Pattern Classification (duda, hart, stork) - A must read for object recognition.

Edit: Important things to know(research): sample size, test set vs training set, you need a "Basis" in order to extract.

edit flag offensive delete link more

answered 2011-10-05 19:52:27 -0600

Object recognition is a very hard problem, and there is still a lot of research going on. That said, you might be interested in Bastian Steder's work on a 3D Point Cloud Based Object Recognition System.

edit flag offensive delete link more

Question Tools

1 follower


Asked: 2011-10-05 17:22:56 -0600

Seen: 4,137 times

Last updated: Oct 06 '11