Variable precision sensor fusion: An evidential classification approach

This research addresses the problem of fusing information from disparate sensors to support decisionmaking in an intelligent system. The sensors are members of a sensor suite which are integrated to form a multisensor system. The information produced at each sensor represents its characterization of a highly complex, unstructured and uncertain environment. Some sensors observe disparate environmental features and their observations are likely to be made at varying levels of precision.; Both numerical and symbolic methods are used to fuse and classify sensory output. Typical of the numeric or quantitative methods is the use of Bayesian analysis and statistical methods to fuse observations (Chair and Varshney, 1986) and statistical pattern recognition to classify the results (Duda and Hart, 1973 and Van Trees 1969). Symbolic, or qualitative methods generally consist of extensive rule sets designed to operate on the combined output from the sensors. The application of the rule set to the fused output is then the classification process (Benoit and Laskowski, 1988).; The methodologies presented in this research combine the quantitative representation of belief and the qualitative representation of knowledge. The fundamental thesis is to demonstrate that quantitative methods for assessing belief can be combined with more symbolic methods for representing knowlege to produce an effective evidential classification process. The interaction between the quantitative representation of the combined observations and the symbolic knowledge base constitutes the evidential classification process. There are two fundamental objectives of this research: (1) to develop a methodology for fusing information observed at the sensors using a mix of Bayesian and variable precision evidential reasoning techniques, and (2) to derive a classification methodology which will operate on the evidential support levels used to represent the fused observations.; The major contribution of this research is the development of a mathematical technique to combine information in the form of Bayesian belief estimates, coarsen these estimates to belief functions at an appropriate level of precision, based on impaired sensor performance and/or occlusions which might exist, and finally classify the resulting evidence by allowing it to interact with a knowledge base of aggregate belief functions.