Acoustic traffic classification using an artificial neural network

Abstract

Traffic noise and/or community noise can be measured with an unmanned measurement station which continuously records the sound pressure level (e.g. Symphonie measurement system). If wanted or needed, the measurement equipment can be configured to record all sounds exceeding a previously defined trigger level. For labeling or classification of the source type, from which the recorded sound originates, the recording must be listened to and manually classified. The desire to render this classification less time consuming suggests the development of an automatic method for sound source classification. In this thesis, the development of such a method is aimed at. The choice of an Artificial Neural Network as a classifier is motivated by its design model; the human brain and nervous system, and furthermore; the human ability to accurately distinguish different sounds. Sounds from heavy and light traffic (e.g. trucks and cars respectively) have been recorded, preprocessed and successfully classified. The preprocessing techniques used are filtering, resampling, signal modeling (ARMA-model) and Principal Components Analysis. The Neural Network employed for source type selection is a Multi Layer Perceptron with one hidden layer. One key issue is the extraction of features which defines and separates the different source types. Method performance is validated by simulation of new measurements and classification thereof. The results show that the classification is 94 % accurate for the specific measurement situation. For assessment purposes, the performances of two reference methods are compared with the artificial classification. Manual classification of the recorded sounds was 96 % accurate and a method utilising the euclidean distance from new, unknown vehicles to the class average in feature space was 83 % accurate.