Emergency Vehicle Detection in Noisy Environments Based on Feature Fusion

To enhance automotive safety, in this paper, a method is studied to accurately and efficiently detect emergency vehicles from the siren sounds they emit in the road environment and alert the driver. The resultant novel vehicle-mounted siren detection approach is proposed that leverages the fusion of spectral-temporal domain features. It explores the complementary nature of acoustic features between the frequency and temporal domains. A fast Fourier transform is conducted on the input sound signal to derive its frequency domain representation, followed by the computation of the log-Mel spectrogram using a log-Mel filter bank. In parallel, a convolutional neural network is employed to model the sound waveform directly in the temporal domain, yielding the temporal representation of the signal. The log-Mel spectrogram and temporal representation are integrated to serve as the input for the classification network. Our experimental results demonstrate that the proposed method achieves a detection accuracy of 97.58% on the LSAD-EVSRN dataset, surpassing the performance of methods that rely solely on frequency or temporal domain information. This finding highlights the effectiveness of the feature fusion technique in significantly enhancing the model's detection capabilities.