IEEE Access (Jan 2024)
Smart City Traffic Management: Acoustic-Based Vehicle Detection Using Stacking-Based Ensemble Deep Learning Approach
Abstract
Acoustic data analysis has emerged as a critical area of exploration for the detection of different events for quick actions in smart traffic management systems, particularly in traffic management and safety as a step toward smart cities. A specific challenge is to precisely classify road noises and emergency vehicles using sound, which is essential for speeding up emergency response times and improving traffic flow management. While existing solutions address this problem, there is opportunity for enhancement in terms of precision and accuracy to enhance the traffic flow in a sustainable smart city through a demanding and innovative technique. In this study, we suggest stacking ensemble deep learning techniques to intelligently classify emergency vehicle sirens from various background noises using a data set of traffic collected on roads via microphone sensors. The ensemble model incorporates Multi-Layer Perceptron (MLP) and Deep Neural Network (DNN) as base-learners, with an LSTM model as a meta-learner. This approach not only optimizes model efficiency but also facilitates advanced feature engineering to extract useful features including Mel Frequency Cepstral Coefficients (MFCC), Z-score, root mean square (RMS), spectral centroids, spectral flux, mel spectrogram, chroma, contrast, and Tonnetz. Using these features, our proposed Stacking Ensemble LSTM successfully classified traffic noises and emergency vehicle sirens with the highest efficiency. Upon evaluation of the test set of data for our proposed model, it has gained an accuracy of 99.12% with F1 scores ranging from 98%. This significant improvement highlights the dominance of our proposed model approach over prior research. Our proposed model presents assurance in advancing traffic control and safety statutes, demonstrating potential applicability in daily intelligent transportation systems.
Keywords
