Markov chain prediction fusion for automatic target recognition

We introduce the temporal target recognition problem, in which information is aggregated over time. The simplest data fusion approach (multiplication of class conditional probabilities) is shown to give poor results when the sequence of information obtained is not independent. We describe a novel algorithm which models target behaviour as a Markov process, with a simple distribution model within each state being used to quantify the degree to which current information is independent of previous information. This new fusion algorithm, which we refer to as the Markov chain prediction fusion technique, is evaluated on realistic artificial data and the experimental results are presented.