Acoustic-based Vehicle Detection for Smart City Traffic Management using Stacking Ensemble Deep Learning

Abstract

In order to ensure timely traffic control and emergency response, efficient mechanisms are needed to be able to distinguish traffic noise from the sirens of emergency vehicles in smart city traffic management systems. In this paper, we present a vehicle detection system based on an acoustic approach and stacking ensemble deep learning. To make the classification of traffic sounds effectively, the proposed model works on the acoustic features including fundamental frequency, formant structure and loudness rather than Mel-Frequency Cepstral Coefficients (MFCCs) and Mel-spectrogram features, which are used in conventional methods. These features are useful for the separation of emergency sirens from the background traffic noise.The stacking ensemble architecture stacks the benefits of a few deep learning models to enhance classification performance. The experimental results show that the GRU enhanced stacked model can obtain an accuracy of 92%, precision of 90%, recall of 89% and F1-score of 89% which are better than the individual baseline models.