Author: Clemens Kubach
This repository is one of three for my bachelor thesis on "Development of an Embedded System for Detecting Acoustic Alert Signals of Cyclists Using Neural Networks".
It contains the architectures of the used neural networks to detect the sound event of cyclists using their bicycle bell. Written in Python 3.9.
The other related repositories are:
You can use this package with:
pip install git+https://github.com/ClemensKubach/bicycle-bell-sed-models.git
If this repository is private, use pip install git+https://{gh_token}@github.com/ClemensKubach/bicycle-bell-sed-models.git
where {gh_token} is your personal access token to your github account with the rights to clone private repositories. You must have granted permissions to access this repository from GitHub.
There are 3 model configurations...
A model based on a CRNN architecture without any pre-training. Inspired by [1].
The pre-trained YAMNet[2][3] base model without any transfer learning. The resulting probability values for the class "Bicycle bell" is directly taken out of the results of all 521 classes and the maximum probability from the windows of the wave input file is taken.
An extended pre-trained YAMNet model with transfer learning using embeddings of the base model[2][3]. A classifier with LSTM and dense layers are followed.
[1] Lim, H., Park, J., & Han, Y. (2017, November). Rare sound event detection using 1D convolutional recurrent neural networks. In Proceedings of the detection and classification of acoustic scenes and events 2017 workshop (pp. 80-84).
[2] https://tfhub.dev/google/yamnet/1
[3] https://github.com/tensorflow/models/tree/master/research/audioset/yamnet