The Skeptical Explorer: A Multiple-Hypothesis Approach to Visual Modeling and Exploration

Abstract

The primary intent of this work is to present a method for sequentially associating three-dimensional surface measurements acquired by an autonomous exploration agent with models that describe those surfaces. Traditional multiple-viewpoint registration approaches are concerned only with finding the transformation that maps data points to a chosen global frame. Given a parts-based object representation, and assuming that the view correspondence can be found, the problem of associating the registered data with the correct part models still needs to be solved. While traditional approaches are content to group segmented data sets that geometrically overlap one another with the same part, there are cases where this causes ambiguous situations. This paper addresses the model-data association problem as it applies to three-dimensional dynamic object modeling. By tracking the state of part models across subsequent views, we wish to identify possible events that explain model-data association ambiguities and represent them in a Bayesian framework. The model-data association problem is therefore relaxed to allow multiple interpretations of the object's structure, each being assigned a probability. Rather than making a decision at every iteration about an ambiguous mapping, we look to the future for the information needed to disambiguate it. Experimental results are presented to illustrate the effectiveness of the approach.