Evidence-based Regression Trees for Object and Event Classification

摘要

Existing approaches to object and event classification have focused on data fusion requiring a priori probability distributions. The hierarchical nature of object (buildings, bridges, etc) development in two dimensional imagery readily lends itself to a more rigorous mathematical modeling with prediction capabilities. This paper presents a new approach to the classification of hierarchical (time and space) object classification based upon an innovative Evidence-based Regression Tree (ERT) prediction mechanism. Typical regression trees, based on classical Bayesian probability theory, require the specification of a complete probabilistic model, and assume certain knowledge of all prior probabilities. Our ERT approach is based on logistical constraints founded on evidence and uncertainty principles; it exploits probability intervals measuring both the level of belief and the amount of information extracted from imager sources. We believe our system will facilitate the production of anticipatory intelligence from diverse information sources into the new products needed for the changing threat environment.