Acoustic Traffic Attribute Classification for ITS: A Comparative Study of Machine Learning and CNN Approaches

Reliable information about traffic attributes, including vehicle type, speed range, and direction, is essential for traffic management and intelligent transportation systems (ITS). Although radar- and camera-based solutions can provide accurate data, they are often expensive, vulnerable to environmental conditions, and may raise privacy concerns.In this study, we investigate passive acoustic sensing as a cost-effective and privacy-preserving alternative. Using stereo vehicle pass-by noise encoded as Mel-Frequency Cepstral Coefficients (MFCCs), we simultaneously classify vehicle type, speed range (inferred from road-specific speed limits), and movement direction. Three modeling strategies are evaluated: (1) traditional machine learning on time-averaged MFCCs, (2) a hybrid model combining ResNet50-based feature extraction with LightGBM classification, and (3) end-to-end convolutional neural networks (CNNs), including a lightweight multi-task variant enhanced with attention mechanisms.Experiments are conducted on the public IDMT-Traffic dataset, comprising 17,506 stereo audio clips. Our multi-task CNN achieves the best overall performance with only 646K parameters, reaching approximately 99% accuracy across all three tasks. ... mehrThese results highlight the viability of stereo acoustic sensing as a lightweight, scalable, and privacy-preserving solution for real-time traffic perception and V2X infrastructure support.