Blind separation of multiple vehicle signatures in frequency domain

This paper considers the problem of classifying ground vehicles using their acoustic signatures recorded by unattended passive acoustic sensors. Using these sensors, acoustic signatures of a wide variety of sources such as trucks, tanks, personnel, and airborne targets can be recorded. Additionally, interference sources such as wind noise and ambient noise are typically present. The proposed approach in this paper relies on the blind source separation of the recorded signatures of various sources. Two different frequency domain source separation methods have been employed to separate the vehicle signatures that overlap both spectrally and temporally. These methods rely on the frequency domain extension of the independent component analysis (ICA) method and a joint diagonalization of the time varying spectra. Spectral and temporal-dependent features are then extracted from the separated sources using a new feature extraction method and subsequently used for target classification using a. three-layer neural network. The performance of the developed algorithms are demonstrated on a subset of a real acoustic signature database acquired from the US Army TACOM-ARDEC, Picatinny Arsenal, NJ.