Implementation of a neural network on a microcontroller for recognition of warning signals

The development of neural networks is expanding onto platforms with a lesser computational power such as microcontrollers. As a result of this development, solutions from neural networks can be compressed and implemented on everyday products. The microcontrollers physical footprint is relatively small compared to computers can be used in order to create "smart" products, which take in and process sensordata from the surrounding, possible in combination with a neural network. In this thesis, a deep neural network (DNN) was implemented on an STM32F746NG microcontroller unit in order to primarily recognize the Yaris car horn. The network classified the car horn with adequate accuracy and with a latency of 120ms. This was a result of experiments in order to evaluate four different neural networks (deep neural network, convolutional neural network, convolutional recurrent neural network, depthwise separable convolutinal neural network). The networks were trained on a computer with a data set created during the project and implemented on the microcontroller unit