A comparison of analysis methods for vehicle classification by laser vibrometry

In this report laser vibrometry data from six different vehicles are analysed by four different analysis methods. The vehicles are the track-laying vehicles Bv206, Strf90, T72, Strv121, and on wheels Tgb11 and Tglb30. The data is collected at tree different test sites, and in various conditions. The frequency modulated data is first preprocessed by peak detection, and transient reduction. Features are extracted from power spectral density spectra (PSD), autoregressive model parameters (AR), Morlet wavelet spectra, and by empirical mode decomposition (EMD), and Hilbert spectra. Six elements are used in each feature vector. The feature vectors of each vehicle are divided in reference data and test data. The test data is classified by Mahalanobis classification and associated to the nearest reference data class. All together 222 measurements are used. Best result is achieved by the EMD-method and 62% of the test signals are assigned to the right reference class in the six class case, without regard to differences in the engine rpm or surfaces illuminated. Feature vectors of dimension five are also classified. The best result is again achieved by the EMD-method but here only 56% of the test signals are assigned to the right reference class in the six class case. The focus here is primarily on the comparison of the analysis methods and it is suggested that a higher classification percentage could be achieved by testing one feature vector at a time leaving the rest to a better estimation of the mean and the covariance matrix of the reference class. Feature vector element values are not included in the appendices of the report but are available as a supplement.