Smart Urban Traffic Control: Acoustic Vehicle Detection Utilizing Stacking Ensemble Deep Learning

Abstract

Improving urban mobility in the context of smart city traffic management necessitates efficient monitoring and quick emergency response. This work extends the existing stacking-based ensemble deep learning approach for acoustic-based vehicle detection with a focus on improving classification accuracy for road noise and emergency vehicle sirens, such those of ambulances. While the original approach layered Fully Connected MLP, Deep Neural Network (DNN), and LSTM networks, recent study adds a lightweight Gated Recurrent Unit (GRU) layer to further increase performance. To optimize handling of large datasets and further improve accuracy, the LSTM classifier is paired with GRU, which is renowned for its ability to boost processing speed and prediction accuracy by simplifying the design of recurrent neural networks. With an incredible 100% accuracy, the extended model's LSTM and GRU combo performs better than previous configurations. This model uses acoustic properties including fundamental frequency, loudness, and amplitude to more precisely identify road noise and emergency vehicle noises. The proposed technique outperforms traditional algorithms that just employ Mel Frequency Cepstral Coefficients (MFCC) or Mel spectrograms, offering a fresh approach to real-time traffic management and emergency vehicle priority in smart cities. The performance of all models, including the extended LSTM-GRU technique, is evaluated using accuracy, precision, recall, and F-score metrics, demonstrating the efficacy of the proposed system for future traffic management in cities