Congest-Net: A new CNN model for Audio based Traffic Congestion

Nowadays, traffic congestion is continuously increasing, costing time lost, affecting citizen’s quality of life, and increasing gas emissions, exacerbating pollution problems. Many sensed modalities can be analysed for detecting traffic parameters and congestion state. Nevertheless, the audio modality has not been sufficiently exploited, as the other traffic signals, although it seems to be much attractive for its lowest cost and its intuitive use in our daily life in congestion detection. In this paper, we propose an efficient system for audio traffic congestion classification, based on spectrogram generation and a new CNN architecture, called Congest-Net. Based on public real life traffic videos, three traffic flow states were considered, namely the Light, the Medium and the Heavy Flows. Experimental validation demonstrates the efficiency of the proposed system, with performances in term of mean accuracy and mean precision of $\mathbf{96.42 \%}$ and $\mathbf{96.41 \%}$ respectively.