Deep Learning Stack Ensemble Approach to Acoustic-Based Vehicle Detection of Smart City Traffic

In smart city traffic environments, distinguishing emergency sirens from background road noise is critical for effective traffic management. This paper presents an acoustic-based vehicle detection system using a novel stacking ensemble deep learning model that combines Multi-Layer Perceptrons (MLPs), Deep Neural Networks (DNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRUs). Unlike conventional methods relying on MFCC or Mel spectrogram features, our approach utilizes raw acoustic features—such as fundamental frequency, formant structure, and loudness—to improve robustness against noise. The integration of a GRU layer enhances processing speed and optimizes sequential modeling. Experimental results demonstrate that the GRU-enhanced ensemble model achieves superior accuracy (92%), precision (90%), recall (89%), and F1-score (89%), making it highly effective for real-time emergency siren detection in smart city scenarios.