Acoustic Traffic Event Detection in Long Tunnels Using Fast Binary Spectral Features

In this paper, we study the traffic event detection from audio signals. Real-life data are collected in a long tunnel, and audio samples are labeled in accordance with traffic events including tire friction sound, vehicle percussion sound and other background sounds. Efficient spectral features are proposed for the fast classification of audio events. In order to model the acoustic characters, deep neural network approach is adopted. Several state-of-the-art algorithms are used for comparison, including LSTM neural network and Gaussian mixture models with Mel frequency cepstral coefficients. A novel convolutional neural network architecture which processes the input audio data in an end-to-end fashion is adopted for our traffic event detection application. Furthermore, we use time delay estimation algorithms to locate the sound location when the incident happens in the long tunnel. By comparison with the state-of-the-art audio detection methods, our proposed efficient spectral features are proved to be more accurate and more efficient in the detection of audio events related to traffic incidents.