Application of autoencoder to traffic noise analysis
Andrzej Czyżewski, Adam Kurowski, Szymon Zaporowski · Proceedings of meetings on acoustics · 2019
The aim of an autoencoder neural network is to transform the input data into a lower-dimensional code and then to reconstruct the output from this representation. Applications of autoencoders to classifying sound events in the road traffic have not been found in the literature. The presented research aims to determine whether such an unsupervised learning method may be used for deploying classification algorithms applied to the automatic annotation of road traffic-related events based on noise analysis. Two-dimensional representation of traffic sounds based on 1D convolution was fed the core of autoencoder neural network, and after that classified with seven feed-forward classification subnetworks. Obtained results show that sound recordings can help determine the number of vehicles passing on the road. However, instead of being treated as independent, this method output should be combined with another source of data, e.g., video processing results or microwave radar data readings. Results of vehicle types classification and occupied lane obtained with the use of autoencoder are shown in the paper.