Acoustic target classification using multiscale methods

This study considers the classification of acoustic sig-natures using features extracted at multiple scales from hierarchical models and a wavelet transform. In the model-based approach, multiscale spectral fea-tures are extracted with hierarchical autoregressive and moving average (ARMA) models. The model-ing approach is also used for monitoring vehicular activities from an AR spectrogram. The AR spec-trogram shows engine speed, gear changes, and other vehicular activities well, because it represents domi-nant spectral peaks better than a short-time Fourier transform. In the wavelet-transform based approach, multiscale features are obtained with a wavelet trans-form. Multiscale classification methods were applied to acoustic data collected at different test tracks un-der various testing conditions. In this experiment, about 92 percent of vehicles were correctly identified