Date of Award
MS in Electrical Engineering
A Study of Dempster-Shafer’s Theory of Evidence in Comparison to Classical Probability Combination
Scott J. Seims
This thesis is an assessment on the effectiveness of Dempster-Shafer’s Theory of Evidence in comparison to Classical Probabilistic Combination as it applies to Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR). A three feature based classifier (peaks, corners and edges) ATR system is presented. These classifiers are assumed to be independent. The results of both the weight-based Maximum Likelihood and Dempster-Shafer’s Theory of Evidence data fusion techniques are presented.
Using Dempster-Shafer an accuracy of 77.50 percent is obtained, which is less than the 86.25 percent accuracy of target-specific weight-based Maximum Likelihood. Inagaki’s Unified Combination Rule is implemented as a means to increase SAR ATR accuracy and explore further modifications of Dempster-Shafer. Inagaki’s Unified Combination Rule contains Yager’s and Dempster’s Combination Rules as well as Inagaki’s Extra Rule. The maximum accuracy achieved using Inagaki’s Unified Combination Rule was 75.00 percent.
It was concluded that this application lends itself better to classical probabilistic combination. Due to the single sensor (SAR) and the quasi-independence of the three feature based classifiers, Dempster-Shafer’s Theory of Evidence can not be utilized to its full potential.
This thesis is a continuation of Hausdorff Probabilistic Feature Analysis in SAR Image Recognition by Chessa Guilas. Thesis research is directed by Dr. John A. Saghri and sponsored by Raytheon Space & Airborne Systems, El Segundo, California.