Spectrogram-based classification on vehicles with modified loud exhausts via convolutional neural networks

Abstract

Hong Kong is one of the densest cities in the world with busy roads close to residential areas. Excessive noise emitted by vehicles with modified loud exhausts causes further noise annoyance during late night. Existing technologies are inefficient for detecting these undesirable vehicles due to various constrains such as the use of sound level meter requires manual spectrum analysis, and acoustic camera is expensive, bulky for low frequency detection and computationally heavy. In this research, the noise spectral characteristics at the peak sound pressure levels from vehicles observed with and without modified loud exhausts passing-by the measurement points were collected and scientifically analyzed. A spectrogram-based sound recognition prototype with low-cost design was therefore developed in assisting the detection of vehicles with modified loud exhausts. A large amount of audio data was converted into spectrograms by short-time Fourier transform for machine learning and a trained convolutional neural network model based on AlexNet architecture utilized spectrograms as the inputs for classifying whether the pass-by vehicles are with modified loud exhausts. The prototype also deploys YOLO v4 Tiny algorithm for object recognition and is able to distinguish the types of pass-by vehicles i.e. passenger car (which is the target for detection), goods vehicle, and motorcycle for enhancing the usability. The results show that the prototype had reached an overall 96% accuracy in the detection of passenger cars with and without modified loud exhausts.