Acoustic Vehicle Classification Using Mel-Frequency Features with Long Short-Term Memory Neural Networks

Abstract

Monitoring vehicle traffic at a large scale is a challenging task for authorities, particularly considering the high cost of traffic sensors such as vision cameras. To meet the growing demand for more accurate traffic monitoring, the use of traffic sounds has become a popular approach, as it provides insight into the types of traffic present. This paper reports on an approach to vehicle classification based on acoustic signals, using the Mel-Frequency Cepstral Coefficients (MFCC) and the Long Short-Term Memory (LSTM) networks. This study exhibited classification accuracy scores of 82-86.2% across four vehicle categories: motorcycle, car, truck, and no traffic. The results demonstrated that large-scale, low-cost acoustic processing can be effectively used for vehicle monitoring.