Acoustic Vehicle Classification using Deep Learning Trained on a Spectrogram and Scalogram Fusion

Abstract

This paper explores an audio-based on-road vehicle classification method that utilizes visual representations of sound through spectrograms, scalograms, and their fusion as features, classified using a modified VGG16 Convolutional Neural Network (CNN) architecture. The proposed method offers a non-intrusive, potentially less costly, and environmentally adaptable alternative to traditional sensor-based and computer vision techniques. Our results indicate that the fusion of scalogram and spectrogram features provides enhanced accuracy and reliability in distinguishing between vehicle types. Performance metrics such as training and loss, alongside precision and recall of classes, support the efficacy of a richer feature set in improving classification outcomes. The fusion features demonstrate a marked improvement in distinguishing closely related vehicle classes like 'Cars' and 'Trucks'. These findings underline the potential of our approach in refining and expanding vehicle classification systems for intelligent traffic monitoring and management.