Characterising the urban sonic environment is usually achieved by measuring energetic indicators such as the average A-weighted level (LAeq) or the day-evening-night Level(Lden). The European 2002/49/EC directive compels large cities to set up noise maps and implement action plans aimed at reducing the citizens' exposition to excessive noise. Yet the diagnosis step of the process requires the sonic environment to be quantified and the soundscape concept suggests that the nature of sound sources is central in this description. Furthermore, the nature of sources on noise annoyance have been considered homogeneous for city dwellers in some previous studies. Therefore, our work focuses on the nature of sound sources in urban areas and their automatic assignment to given perceptual categories. This is done using very large microphone array processing and video tracking to extract individual passing-by vehicles audio signal from a scene involving multiple vehicles. These moving source signals feed a categorisation process consisting in a supervised machine learning algorithm (Support Vector Machine). The paper presents a benchmark experiment that allowed to gather reference signals out of traffic for different passing-by vehicle types (heavy vehicle, utility truck, personal car, motorcycle, ...) and driving condition (constant speed, acceleration, deceleration). These signals are used as training samples for the machine learning algorithm so that the categorisation process allows to characterize the vehicle on perceptual categories (combining both vehicle types and driving conditions). Another experiment led in situ is also analysed. The automatic recognition robustness is discussed and improved by adding the in situ extracted signals to the training samples.