Target Recognition Fusion Based On Belief Function Theory

Effective processing of uncertain information from each sensor is a fundamental problem in the area of multi-sensor information fusion for target recognition. Belief function theory has been widely used as a significant mathematical model for uncertain information processing. This dissertation investigates the representation and combination of uncertain information in target recognition based on belief function theory.In the process of target recognition based on belief function theory, the first problem is transforming the recognition results of current targets from each sensor to basic belief assignment functions which can be used in belief function theory. In Chapter 3, according to the various forms of information on different level, the methods for basic belief function constructing are designed. On the abstract level, a modification of classical method based on confusion matrices for basic belief assignment functions constructing is suggested. On the rank level, basic belief assignment functions are obtained from the construction of position matrices of classes'indices. On the measurement level, a method based on referential vector, a method based on whitenization weighed function, and a method for consensus support function construction are presented respectively.Dempster rule of combination is the most classical rule in common use for combining pieces of evidence represented by basic belief assignment functions. Because of the counterintuitive results from highly conflicting pieces of evidence combining, the more rational and effective rules of combination are required. The premise of this step is conflict analysis and measurement between pieces of evidence. In Chapter 4, Shafer's conflict measurement and Daniel's idea of potential conflict are incorporated to present a generalized conflict measurement. This measurement of conflict includes both uncommitted basic belief mass in conjunctive combination rule and the incompatible mass from compatible focal elements but not completely overlapping. This measurement is more comprehensive than the conflict measurement in existence, and provides important reference to the construction of combination rule.After deciding the form of conflict measurement, the construction of combination rule of pieces of evidence can achieved according to the value of quantitive conflict. The key step of this process is the proportional redistribution of conflict mass to the combined focal elements. In chapter 5, on the one hand, a synthetic rule of combination is advanced based on Shafer's conflict measurement. Except the rational conflict mass redistributing and focusing downwards weighting, this rule is commutative, and can preserve the neutral impact and conditioning impact of corresponding pieces of evidence. On the other hand, on the basis of generalized conflict measurement, a new rule of combination is advanced. This rule can solve the problem of conflict and focusing downwards together, and can design an algorithm for the rule to satisfy the quasi-association.For weakening the impact of low reliability sensor and ensuring the robust effect in the process of fusion, the sensor reliability evaluating and evidence discounting are investigated in chapter 6. On the basis of the framework of sensor reliability evaluating in existence, the static factors of the sensors'reliability are obtained from confusion matrices, and then selected by the local decision of each sensor for discounting calculation. The dynamic factors of the sensors'reliability are obtained from the direct conflict degree between the output information of each sensor and the consensus, the corresponding discounting rule based on the idea of proportional conflict redistribution in conflict processing can assign the discounted mass more refined.All the method presented in this dissertation have been verified by the numerical examples and experiments on the data of radar aerial targets, and the advanced methods and conclusions of this research can provide theoretical basis of uncertain information representation and combination in the problem of multi-sensor information fusion for target recognition.