Geoff Taylor, PhD Thesis, Monash University
|
|
Chapter 5
Object Classification and Modelling
The ability to locate and classify a variety of unknown objects (cups, bottles, books, etc) is a crucial skill for a flexible, autonomous humanoid robot. In this work, objects are identified as generic collections of geometric primitives (for example, a can is always a cylinder with two end-planes). Object modelling is supported by a split-and-merge segmentation algorithm that attempts to extract geometric primitives from 3D range data. The detection of large regions of consistent shape is facilitated by a novel, robust, non-parametric surface type classifier proposed in this thesis. The new method distinguishes six surface shapes without the need for arbitrary approximating functions as required by conventional surface type classifiers, and is experimentally shown to achieve greater robustness to noise, which ultimately simplifies segmentation.
Mirror Experiment: Interference Rejection (Section 5.7.1)
The object classification and modelling algorithms presented in this chapter
were testing using experimental range data from our robust
light stripe scanner. The first test scene contained simple domestic objects
(a box, ball and cup) which can be described by simple convex primitives, while
the second scene contained some non-convex objects composed of multiple geometric
primitives. For each scene, VRML models for both the raw colour/range data (sub-sampled
at half and quarter actual scanner resolution)
and the extracted convex primitives are presented below. MPEG "fly-thoughs"
allow each model to be viewed without a VRML browser.
The object classification and modelling results for two additional scenes
can be found in Chapter 7: Integration.
|
| ||
 |
||
| ||
| ||
 |
||
| ||
| ||
 |
||
| ||
| ||
 |
||