Two-stage Acoustic Event Detection and Classification for Horn Signals in Urban Scenarios: Comparing VGGish, YAMNet and Mel Spectrogram Approaches with mRMR Feature Selection

Urban vehicle accidents highlight the need for intelligent systems that can process environmental acoustic signals and improve driver awareness. Traditional acoustic event detection and classification (AEDC) methods often struggle to perform effectively when multiple acoustic events overlap in frequency. This study presents a two-stage framework for detecting and classifying horn signals in urban driving scenarios. The first stage separates horn signals from background noise using Log-Mel spectrogram features and a bagging ensemble classifier. The second stage classifies detected horns into four categories (boat, car, train, truck) using VGGish features. Feature dimensionality is reduced by approximately 40% through minimum Redundancy Maximum Relevance (mRMR) selection without losing accuracy. A post-processing step based on the binomial probability distribution corrects errors across consecutive frames in both stages. The proposed system achieves a minimum F1 score of 0.82 before post-processing and 0.96 after post-processing across all horn classes. This represents more than a 15% improvement over recent deep learning approaches, while maintaining low computational demands through feature selection. The hybrid architecture makes AEDC practical for driver assistance systems in automotive environments with limited processing resources.