Predicting vehicle category using psychoacoustic indicators from road traffic pass-by noise

Abstract

A set of road traffic pass-by noises containing more than 2000 vehicles was recorded following the Statistical Pass-By (SPB) methodology (ISO 11819-1:2022). Besides the acoustic descriptors, psychoacoustic indicators (loudness, sharpness, roughness, fluctuation strength) were retrieved for each pass-by of the three vehicle categories defined in the standard (passenger cars, dual-axles and multi-axles heavy vehicles). A fourth vehicle category, comprised of delivery vans, was also investigated. All psychoacoustic indicators significantly differed among vehicle categories, meaning that not only intensity descriptors but also temporal and spectral features of pass-by noise distinguish those classes. With enough instances and a balanced dataset across groups, a machine-learning classification algorithm was trained with 70% of the dataset to predict vehicle categories using the psychoacoustic indicators. Classification accuracy on the test set reached 72.5%. Accuracy losses were primarily caused by 25% of the actual passenger cars being misclassified as vans and vice-versa. Pooling these two categories increased accuracy to 82%. With more descriptors from road traffic pass-by noise than uniquely its maximum noise level, limiting definitions of vehicle categories may be overcome. As a result, measurements such as the SPB can become broader and vehicle fleets worldwide more consistently represented in terms of noise perception.