Recognition of High-Range-Resolution (HRR) Profile Signatures of Moving Ground Targets for Combat Identification (CID)

The objective of this proposal is to demonstrate the advantages of using a Hierarchical Hidden Markov Model for Aided Target Recognition of High Range Resolution (HRR) radar. A Hidden Markov Model (HMM) based technique has been previously shown to provide aided recognition of HRR with high probability of correct identification and low probability of error. This proposal extends current HMM techniques by utilizing a generalized HMM, known as the Hierarchical Hidden Markov Model, with several attractive properties not found in classic HMMs – in particular superior ability to learn the different stochastic levels and length scales present in the structure of the target features. One key difficulty in the application of any HMM is parameter estimation. The unknown parameters are typically point-estimated in a Maximum A Posterior (MAP) or Maximum Likelihood (ML) sense using an Expectation Maximization algorithm. We propose to utilize a Variational Bayes (VB) algorithm that does not generate a point estimate for the parameters but an approximation to the full posterior of the model parameters. The VB technique has shown in many applications to be less sensitive to overfitting and better-suited for active learning; the VB solution also allows one to perform model selection, here concerning the appropriate number of HMM states.