Acoustic Traffic Monitoring Based on Deep Neural Network Trained by Stereo-Recorded Sound and Sensor Data

In this study, we present a machine-learning-based acoustic traffic monitoring aiming to realize a relatively low- cost system compared with existing traffic sensors. Since vehicles are moving fast and the sound from different vehicles may overlap, the relationship between acoustic signals and traffic conditions, such as the number of vehicles passing or speed, is complicated. Then, the machine-learning approach is attractive. For this purpose, collecting a sufficient amount of data to train, for example, deep neural networks (DNNs), is crucial. In this study, we built a 48-hours dataset using stereo microphones and sensors already installed on highways to label traffic conditions automatically. We used ConvMixer, one of the recently well-used convolutional neural network (CNN) architectures, to estimate four traffic conditions, i.e., the total number of vehicles passing, the number of large vehicles passing, speed, and time occupancy. In experiments, we compare the acoustic features used as input to the DNN, compare our method with conventional methods, and apply our method to traffic flow discrimination.