<title>Bayesian network approach to sensor fusion of aerial imagery</title>

Abstract

Estimation of superresolution models is a problem of great interest across a broad range of applications in computer vision and robot perception. However, approaches to superresolution model estimation tend to have very high computational complexity. In this paper, we address the superresolution model estimation problem using a general modeling approach based on two-layer Bayesian or causal networks. Sensor nodes encode stochastic sensor models, while model nodes encode probabilistic inferences made about their state. The model nodes are arranged as a MRF spatial lattice. We derive optimal estimation procedures for several classes of superresolution world models, including single and multiple observation models, and analyze their computational complexity. We subsequently introduce three suboptimal estimation methods: Reinjection of Marginals (ROM), Independent Opinion Pool (IOP), and Non-Propagation of Neighbors (NPN). These methods, although suboptimal, are extremely efficient and provide high-quality superresolution estimates. We conclude by presenting results from the application of these procedures to the fusion of multiple aerial images to form highly accurate superresolution images for airborne surveying and monitoring applications.