This research project examines some of the basic high level computer vision problems associated with the task of manipulating and assembling structural frames under autonomous visual control. The approach treats high level vision as uncertain inference. Uncertainties arising from the sensing process and scene itself are explicitly represented and resolved with prior knowledge at a high level of abstraction. The uncertain inference process are realized as a parallel network, based on a Markov Random Field formalism. The structure assembly domain focuses attention on geometric reasoning about spatial relationships between parts, particularly in the presence of uncertainty. Finally, the assembly task provides a functional goal that can define criteria for deciding when too much uncertainty is present. This provides a rationale for planning and acquiring new images in an active framework. This research could have a significant impact on the flexibility of future computer vision systems, and may also increase our understanding of human high level vision.
