In-car violence detection based on the audio signal

Abstract

When it is intended to detect violence in the car, audio, speech processing, music, and ambient sound are some of the main points of this problem since it is necessary to find the similarities and differences between these domains. The recent increase in interest in deep learning has allowed practical applications in many areas of signal processing, often surpassing traditional signal processing on a large scale. This paper presents a comparative study of state-of-the-art deep learning architectures applied for inside car violence detection based only on the audio signal. The methodology proposed for audio signal representation was Mel-spectrogram, after an in-depth review of the literature. We build an In-Car video dataset in the experiments and apply four different deep learning architectures to solve the classification problem. The results have shown that the ResNet-18 model presents the best accuracy results on the test set.