Detecting emergency vehicles in traffic and developing such vehicle-aware siren systems are important to provide the right traffic flow. Nowadays, autonomous vehicles are widespread. These vehicles provide a safe drive by using data obtained via different sensors. The sound sensors, which are one of those sensors, are significant to warn drivers in an emergency or allow emergency vehicles to be the priority for a transition. In this study, a model has been tried to be learned, which distinguishes emergency vehicle sounds from traffic sounds by using Convolutional Neural Networks (CNN). An artificial neural network model has been obtained by using the Keras Deep Learning Library. ReLU has been preferred as the activation function and the Adam optimization algorithm has been used, which updates the learning rate in real time for each parameter. A labeled sub-dataset has been constituted of 314 sound files, containing traffic, car horn, ambulance, police and fire truck siren records from the AudioSet dataset. Mel Frequency Kepstrum Coefficients (MFCC) values obtained from the sound data using the Librosa library have been given as input to the deep learning model. The labeled sub-dataset has been divided into a training set and a test set as 80% and 20%, respectively. The epoch parameter of the neural network has been set as 200. After the training, emergency vehicle siren sounds are classified with an 80% accuracy rate.
Anahtar Kelimeler: Machine Learning, Deep Learning, Convolutional Neural Networks, Signal Processing, Emergency Vehicle Sirens
|
