Detection of Moving Vehicles From Multichannel Acoustic Signals Using Convolutional Recurrent Neural Networks

This article presents a method for detecting the acoustic signals of moving vehicle in a field environment using a convolutional recurrent neural network (CRNN) and multichannel audio. First, to improve the discrimination ability of the proposed method, a sound event detection (SED) framework is applied to detect the moving vehicle, apart from other seven sound event classes which are common in the field environment. Second, the asymmetric convolution block (ACBlock) structure is introduced in the convolution stage of the framework to improve the ability of the convolutional layer to capture the temporal derivative and the characteristics of the feature map in the frequency domain. Then, a channel time-frequency attention (CTFA) module is employed after the convolutional stage to enable the network to prioritize critical information and ignore details irrelevant to the feature map detection task. Finally, the proposed method is evaluated on a dataset generated from audio clips covering different recording situations in the field environment with a total length of 48 h. Ablation studies are performed to demonstrate the effectiveness of the ACBlock and the CTFA module. Comparison with existing methods shows that the proposed method can achieve better performance by showing an improvement of ${F}1$ -score by 8.9%, area under the curve (AUC) by 0.0372, and a reduction of error rate by 5.3%.