Automatic Target Recognition

We propose a new approach to automatic target recognition based on the concept of stochastically deformable templates. In contrast to conventional static approaches, our concept is based on active image tracking. It allows for simultaneous image processing, parameter estimation, and filtering. High target signature variability captured by optical, infrared, millimeter-wave and microwave radar acoustic sensors, etc. substantially diminishes the effective- ness of these technologies in automatic target recognition systems. In this proposal, we address related problems of effective representation and processing of target signatures subject to random fluctuations. Our approach is to represent randomly perturbed dynamically evolving images using the emerging mathe- matical technique of wavelet descriptors for stochastically deformable templates. The wavelet descriptor has many desirable properties including multiresolution representation, invariance, uniqueness, stability and spatial localization. Real time processing of randomly evolving target signatures is seriously restricted by limitations on availability of communication channels, bandwidth length, computer power, etc. One way to circumvent these difficulties is the dynamic distribution of resolution, so that the maximum resolution windows are applied to a small number of the image landmarks. To achieve this goal, we propose a new method of template deformation tracking and prediction, based on the nonlinear filtering approach. Specifically, we use a new fast spectral algorithm for nonlinear filtering. By separation of observations and parameters, this algorithm shifts the main bulk of computations off line. The main objective of the project is to demonstrate that the recent breakthroughs in wavelet-based image descriptors and fast nonlinear filtering make feasible the real-time implementation of our approach.