An approach to detection of events occurring in road traffic using autoencoders is presented. Extensions of existing algorithms of acoustic road events detection employing Mel Frequency Cepstral Coefficients combined with classifiers based on k nearest neighbors, Support Vector Machines, and random forests are used. In our research, the acoustic signal gathered from the microphone placed near the road is split into frames and converted into a 2-dimensional form of Mel-cepstrogram. Next, the sequence of mel-cepstrograms is processed by the autoencoder neural network, which assigns a unique embedding to each of processed mel-cepstrograms. The embeddings may be treated as features which can be fed on the input of other machine learning-based classifiers. In our research, we prepared such an autoencoder and compared it with a standard solution of parameterization consisting of averaging MFCC throughout all the analyzed frames. Both types of features were then treated as an input for selected types classifiers. It was found, that parameters derived by the autoencoder neural network may be useful for improving the performance of classifiers in case of problematic classes such as detection of single and multiple vehicles passes.
Authors
Additional information
- Category
- Aktywność konferencyjna
- Type
- publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)
- Language
- angielski
- Publication year
- 2019
