A second-order uncertainty model for target classification using kinematic data

When kinematic data is used for target classification, traditional Bayesian methods often lead to unreasonable results, especially when the target is not maneuvering. We find that this is due to the poor modeling of the uncertain mapping from the target class space to the maneuver feature space. Our proposed second-order uncertainty model gives both a preferable description of the uncertain mapping from the feature space to the class space and a more practical method to calculate the class likelihood under a relaxed dependence assumption. It is also clarified when the classifying features are extracted from a multiple-model filter, the real dynamic mode of a target should be determined to prevent the classifier from being degraded since it is in fact not a stochastic process and cannot be correctly described by multiple models simultaneously. A numerical example demonstrates that a well-formulated Bayesian classifier can produce results as expected.