Hearing unseen risks: An advanced audio-driven approach for detecting potential road risk events

Traditional road monitoring methods, including radar, light detection and ranging (LiDAR), and video surveillance, face challenges in accurately and comprehensively identifying unexpected events and potential risks owing to high costs, low accuracy, and environmental complexities. This study proposes an audio-based monitoring method for detecting abnormal road events aimed at uncovering hidden traffic risks. Audio samples collected from various sources, such as FreeSound and Zapslat, are preprocessed using noise reduction, detection, integration, and digital signal conversion via sample quantization. An audio recognition classification model based on the RUSBoost algorithm is constructed and optimized using sequential model-based optimization (SMBO). This optimization addresses issues related to imbalanced samples and features. Optimization increased the accuracy from 68.9% to 96.5% and reduced the error rate to 3.5%. The true positive rate (TPR), true negative rate (TNR), and positive predictive value (PPV) also showed significant improvements, along with notable enhancements in the false positive rate (FPR), F1 score, matthews correlation coefficient (MCC), and kappa (KAP). For key audio events such as slides and crashes, TPR and PPV exceeded 93.2%. The SMBO-RUSBoost model distinguished traffic events from noise by extracting multidimensional features from audio data corresponding to various traffic events. Its classification performance provides a solid foundation for traffic risk assessment and decision-making. The model offers a novel approach for improving road safety.