A Deep Wavelet-Fourier Method for Monaural Vehicle Speed Estimation in Real-Life Settings

Vehicle speed estimation using acoustic data offers a cost-effective and non-intrusive approach to enhance traffic safety and efficiency. Previous studies are limited by controlled settings, constraints on vehicle types, and reliance on manual tuning. This paper proposes an end-to-end framework for monaural settings in a multi-lane roadway, subject to real-life ambient noise without restriction on vehicle types. First, we develop a multimodal feature vector by processing raw audio data using a hybrid Fourier-Wavelet method. Second, a careful examination of the ambient noise and vehicular audio guarantees the validity of the suggested feature vector. Third, we design two deep neural networks to handle both regression and classification tasks. Our method is evaluated against state-of-the-art on a benchmark dataset comprising 304 samples. The results demonstrate a substantial improvement in accuracy, increasing by 29.4% (achieving 83.26% accuracy for the target class), and enhancement in the root mean square error (RMSE) for the regression task by 5.6%. In addition, we provide a proprietary dataset collected and curated as part of this study in four urban locations in Melbourne, Australia. This dataset represents the first real-world compilation of complex acoustic speed data, comprising 364 samples. Subsequent tests yield a classification accuracy of 84.02%, as well as Mean Average Error (MAE) and Root Mean Square Error (RMSE) values of 6.38 and 8.09 for regression, respectively.