Next: Multiple View Integration Up: Active Perception Previous: On Integrating Local

Building Volumetric Models from Sensor Data

Authors: [tex2html_wrap4166]A. Lejeune, D. Baird, F.P. Ferrie, P. Whaite

Investigator username: ferrie

Category: perception

Subcategory: active perception

In order to perform tasks such as describing, manipulating, and avoiding collisions with objects, one needs to be able to extract basic information about their three-dimensional size and shape from available sensor data. But these data provide only indirect information about 3-D shape, e.g. how surfaces reflect light or estimated point samples from a surface. Our related work in visual reconstruction deals explicitly with the problem of inferring 3-D surfaces from data provided by television cameras and laser rangefinders. The emphasis of this research, on the other hand, is the problem of interpreting the shape of an object that is represented by a particular surface.

For the purposes of our work the shape of an object is represented by a collection of volumetric primitives, where each describes the coarse geometric characteristics of a particular part. The intent is to be able to generate a unique description of an object in a bottom-up fashion using only general constraints about a particular domain of objects. This kind of description is in itself adequate for many tasks involving object manipulation in a robotics environment. However, a longer term interest is to investigate how such descriptions can lead to generic forms of object recognition.

The formal basis of this work is differential geometry, which contributes to the modeling problem in two related ways. First it provides a mathematical basis for characterizing surface features related to object/ground separation and parts decomposition. In another related project we are investigating the structure of these local features or ``trace points'', particularly how they give rise to contours that separate a surface into different parts. Second, concepts from differential geometry in the large are used to infer the geometric structure of surfaces corresponding to parts, or equivalently, to choose a volumetric primitive that best describes a surface from a finite repertoire.

We have developed algorithms for parts decomposition and the inference of volumetric primitives from surface data. These have been successfully applied to objects whose parts decomposition can be characterized by a dense covering of trace points. However, the general case is more complex and involves cues that must themselves by inferred (similar to the case of subjective contours). We are continuing in this direction, but are also looking at applications in machine vision to gain insight into the general problem.

Next: Multiple View Integration Up: Active Perception Previous: On Integrating Local