Assessing Driving Risk Level: Harnessing Deep Learning Hybrid Model With Intercity Bus Naturalistic Driving Data

Abstract

Driving risk assessment is crucial for enhancing traffic safety, especially given the severe consequences of highway accidents. This study advances the field by developing a deep learning hybrid model for time series analysis to categorize driving risks into low, moderate, and high levels. By collecting naturalistic driving data from intercity buses, the model is trained on an extensive dataset featuring 27,057 journey-based instances, incorporating dynamic GPS data and static journey background information from over 300 drivers. The model’s effectiveness is highlighted by its outstanding weighted average F1-score of 0.932, indicating exceptional robustness and predictive accuracy. Through comprehensive feature engineering and examinations of three temporal neural models, this research identifies the best parameter configurations. The key finding is that including static journey background information leads to improvements by 8.8% on average in model performance. Additionally, the high-risk level prediction F1-score reaches 0.728 for the proposed model, which is up to 9.3 times better than the performance of the machine learning baseline model. This breakthrough in driving risk prediction not only represents a major advancement in traffic safety management but also has practical implications for fleet scheduling management among transportation companies in the future. By applying this model, companies can enhance passenger safety and comfort, showcasing the significant potential of deep learning in real-world applications.