Blind separation to improve classification of traffic noise

In order to apply noise mapping to traffic noise prediction, a knowledge of several information about traffic characteristics is required to predict the noise levels emitted by the roads involved. In the European case, the CNOSSOS-EU calculation method for traffic-noise level prediction is now under discussion, to be agreed in response to the European Directive relating to the Assessment and Management of Environmental Noise (2002/49/EC). In this application context, standard ISO 1996-2:2007 Determination of Environmental Noise Levels, in its Section 6.2, specifically mentions that during Leq measurements of road traffic noise the number of vehicle pass-bys shall be counted during the measurement time interval. This information is often not available in many roads, so it is typically registered by means of casual counts, often through manual procedures. Besides, if the measurement result is converted to other traffic conditions, a categorization of the vehicles involved is also required. Some additional information, such as the traffic density and the average speed, should be registered if a calculation method is used to build a noise map. In this paper a new automatic classification system of traffic noise covering these requirements is presented. The portable system processes a two channel audio recording to provide information of the average speed and the number of vehicles, which are classified in six categories during the measurement period. After several evaluations of the possibilities to get a good classification of the noise emission of a road from audio recordings, it is shown that increasing the within-class separation, as well as introducing a novel BSS–PCA-based classifier, the precision achieved in the final results is substantially improved.