Overview
Experimental research code for the detection of planar objects and articulation models using stereo data.
This package is deprecated. Please look at the articulation stack for the replacing packages. Tutorials are available in articulation_tutorials and a demo using Microsoft Kinect is available in articulation_perception.
Video
Project: 6D poses of planar surfaces from point clouds
Jürgen Sturm
Autonomous Intelligent Systems Lab, University of Freiburg, Germany
Description:
During my stay at Willow Garage this summer, I am focusing on object pose registration from point clouds. In particular, I want to be able to detect, register and track rectangular surfaces that are common in indoor (kitchen) environments, such as cabinet drawers and doors. These pose and shape estimates could then be used to learn articulation models for the interior of a kitchen, as shown in my previous work.
While 6D pose registration is a challenging problem for machine perception in the general case, it becomes (more) feasible when additional assumptions are being made. The stereo cameras of PR2 use structured light to improve the matching process. As a result, they produce highly accurate and dense depth maps. In contrast to 3D laser scanners, the point clouds can be obtained at a normal video frame rates.
We first apply a RANSAC-based approach for segmenting out individual planes in the depth image. Then, for each individual plane, we start at a random start pixel with a rectangle of an initial size, and optimize iteratively its pose and size using a greedy search.
More information (including data sets and videos)
