Robust ground vehicle classification using the modulation spectrogram

Abstract

The performance of a vehicle classification system using acoustic emissions is affected by acoustic interferences such as wind noise. In this paper, we propose using the modulation spectrogam as features for classification. This is done by extraction of the temporal structure in the vehicular sounds that appear as amplitude modulations in the frequency band corresponding to the harmonic structure present in the vehicular sounds. The amplitude modulations are processed in subband channels following critical band frequency analysis. Our results of the visual displays of vehicle sounds produced with the modulation spectrogram are relatively stable in the presence of wind noise. This suggests that the use of the modulation representation is another method for improving the robustness of vehicle classification system. In the paper, we apply the suggested approach to classify between four classes of vehicles and obtain reasonably good classification rate even for data contaminated with interferences.