Abstract:

An in-vehicle detection method was proposed based on the fusion of spectral and temporal features to detect the external emergency vehicle sirens during high-speed driving. The input audio signal was transformed using the fast Fourier transform, and its log-Mel spectrogram was computed to extract spectral features. A convolutional neural network was used to model the raw waveform in the time domain, yielding temporal features. A coordinate attention mechanism was used to fuse and enhance the spectral and the temporal representations. The fused features were subsequently fed into a classifier for final detection. The experiments were conducted on both public and real-recorded datasets. The results show that on the LSAD-EVSRN dataset, the proposed method achieves an AUC (area under the receiver operating characteristic curve) score of 98.92%, with representing an improvement of 14.88% compared to using temporal features alone, and 2.52% compared to using spectral features alone. These results confirm the effectiveness of the fusion strategy, with a high robustness particularly under noisy conditions.

Key words: automotive safety, siren detection, emergency vehicles, sound event detection, feature fusion