| In recent years, multi-sensor data fusion has received significant attention for both military and civilian applications. The studies of multi-sensor information fusion are usually carried out in three different hierarchical levels---data level, feature level, and decision level. Lower level information fusion schemes have the advantage of better fusion performances, but are less robust to the corruption and statistical uncertainty of the data compared with higher level fusion schemes.;In this dissertation, we studied the distributed decision fusion in signal detection. Research in the distributed decision fusion has been focused on the design and implementation of different fusion processors and their applications. In most cases it is assumed that the statistical knowledge of the local decisions is known for implementing the optimal fusion processor, either in the independent or correlated case, and the optimal fusion performance is often reported for ideal scenarios. One problem that happens often in the correlated decision fusion is that the joint statistics of the local decisions are not completely known. Instead of following the popular approach to estimate the joint statistical information between the local decisions, we proposed using the Independent Decision Fusion Processor, which always assumes statistical independence between the local decisions, to fuse the correlated decisions. We derived the conditions under which the Independent Decision Fusion Processor can reach the same performance as that of the Optimal Decision Fusion Processor when dealing with the correlated data.;We also applied the distributed decision fusion scheme to the mine detection problem, which fuses the GPR (Ground Penetrating Radar) and MD (Metal Detector) data. In this application where the correlation information of the local decision data is unknown, the idea of applying Independent Decision Fusion Processor to correlated data was used. A new robust algorithm was developed which addresses a problem in real applications that the sample data used to estimate the statistical information is not representative of the real test data. More discussions on the robustness of the distributed decision fusion were also presented.;Fusion performance comparison was made between the independent decision fusion scheme and the independent data (local likelihood ratio) fusion scheme using a Gaussian log-likelihood ratio model. The Bayesian likelihood ratio fusion was further studied for the cases where either mismatched correlation information or independence assumption is used for the fusion of correlated data. A physical interpretation was given for the analysis results in order to gain a deeper insight of the distributed information fusion concept. |
