Low-Cost Acoustic Sensing Framework for Multivehicle Detection in Complex Urban Traffic

This letter presents a low-cost acoustic sensing framework for vehicle detection and counting in complex urban environments with significant ambient noise. The proposed system is designed for deployment on roadside sensing nodes equipped with single-channel microphones, enabling scalable and continuous traffic monitoring. Unlike existing approaches that rely on controlled conditions or isolated vehicle recordings, the framework targets real-world urban scenarios with multiple simultaneous vehicles and diverse environmental noise. Acoustic data were collected from 20 urban locations in Melbourne, Australia, covering multi-lane roads and heterogeneous traffic conditions with up to three concurrent vehicles per segment. To enhance sensing robustness, acoustic signals are transformed into log-mel spectrogram representations using a Fourier-based analysis, enabling reliable feature extraction under noisy conditions. A lightweight convolutional neural network is employed for vehicle detection, while a compact spectral feature representation is introduced to estimate the number of vehicles in multi-source acoustic scenes. The proposed sensing framework achieves a detection accuracy of 99.96% and a counting accuracy of 93.81% in real-world environments. These results demonstrate that single-channel acoustic sensors can provide reliable, low-cost, and scalable traffic monitoring in complex urban settings.