Abstract
A novel approach for applying hidden Markov models (HMM) to automatic target recognition (ATR) is proposed. The HMM-ATR captures target and background appearance variability by exploiting flexible statistical models. The method utilizes an unsupervised training procedure to estimate the statistical model parameters. Experiments upon a synthetic aperture radar (SAR) database were performed to test robustness over range of target pose, variation in target to background contrast, and mismatches in training and testing conditions. The results are compared against a template matching approach. The HMM captures target appearance variability well and significantly outperforms template matching in both robustness and flexibility.