Fusion methods for multiple sensor systems with unknown error densities (Q1898158)

The paper deals with the multiple sensor data fusion problem in the general framework of a system with multiple sensors, with unknown characteristics. The problem is generally formulated as a special case of the empirical risk minimization problem. The system is supposed to be available so that training data in the form of readings corresponding to objects of known parameters can be obtained. The aim is to estimate a fusion rule, based on the samples, which minimizes an expected square error. Suppose the probability density of the sensors to be unknown and using a suitable ``training'' algorithm, a solution method, based on the empirical risk minimization method, is proposed. It is shown that a fusion rule close to the ``optimal rule'' can be inferred under the assumptions of finite capacity for the class of fusion rules and boundedness of the errors. Some computational issues about estimating fusion rules are given. Several computational methods are determined. The sufficiency conditions for the composite system (the fuser and the sensors) to be better than the best individual sensor are identified. Based on linear approximation and nearest neighbor rule for the special case of Lipschitz continuous functions with application to feedforward neural networks, are presented. A comparison of relative performance of the composite system and individual sensors is given. The class of linearly separable systems is discussed. Further research directions are formulated.