The analysis of vehicle sounds for recognition

摘要

The problem of classifying vehicles on the basis of the acoustic waveform obtained from them has been approached by calculating various central moments of the short term power spectrum of a sample of the signal. It has been found that classification can be performed using two moment measurements, giving good results with vehicles operating under steady running conditions. With a burst of acceleration included in the sample, however, discrimination becomes much more difficult. In the moment space under consideration, it became evident that the movement of sample points with changes of engine speed was itself characteristic of the vehicle class, and this consideration amongst others suggested that the engine speed (or, in practice, the firing rate of the engine) was an important parameter that needs to be determined automatically.The first attempt at finding the fundamental frequency of the waveform was based on autocorrelation, but this gave very unsatisfactory results. The technique of “cepstrum” analysis, however, is shown to give a reliable indication of the firing rate even when the engine sound is deeply embedded in noise. This is in contrast to results obtained by some earlier workers using this analysis in speech studies.