1D convolutional context-aware architectures for acoustic sensing and recognition of passing vehicle type

Abstract

A network architecture that may be employed to sensing and recognition of a type of vehicle on the basis of audio recordings made in the proximity of a road is proposed in the paper. The analyzed road traffic consists of both passenger cars and heavier vehicles. Excerpts from recordings that do not contain vehicles passing sounds are also taken into account and marked as ones containing silence. The neural network architecture employed for these tasks is a 1D convolutional network. Two types of classifiers are tested: one analyzing only the current audio frame and one analyzing three consecutive audio frames that allow us to take into account the context of the middle frame occurrence. The neural network is trained on datasets derived for four frame lengths, namely 50 ms, 100 ms, 200 ms, and 400 ms. Results of statistical analysis of both network classification accuracy are presented. The context-aware variant of a neural network performed better in a statistically significant manner for three out of four investigated frame lengths.