Visibility-based test scene understanding by real plane search

Abstract

We present an algorithm to modelling the test scene for intelligent robot. In intelligent robotics, to develop compact and reliable vision system components for navigation or human computer interaction is essential. As our approach, we develop the line-based modelling and recognition algorithm based on 3D features from stereo camera images. The basic concept is build real plane features from 3D stereo images from mobile robot to navigate or for human computer interaction in the living room environment. The procedure is, first, given 3D line segments, we set up reference plane using principle component analysis (PCA) for each line pair. Then, we measure the normal distance and line orientation for the remains of 3D segments to the reference plane and define coplanarity. And we initiate visibility test to prune out ambiguous planes from reference planes. After we finish visibility test, we patching these reference planes and define them as real plane candidates using plane sweep algorithm. And finally, try to find model objects from the test scene using iterative closest point (ICP). During the implementation, we also use 3D map information for exact localization. We apply this algorithm to the real images and the result found useful to identify door at the wall. © 2008 Springer Berlin Heidelberg.