Dynamic Bayesian Network Approach to Evaluate Vehicle Driving Risk Based on On-Road Experiment Driving Data

In this work, we utilize a dynamic Bayesian network for an inferential analysis of driving-related risks based on our assessment of real-world driving data. On-road driving tests are carried out to gather and analyze data to identify different evaluation indicators of driving-related risks. These indicators include the distance between vehicles, acceleration, steering entropy, visual distraction duration, visual glance speed, and blink frequency. Moreover, these indicators are processed to build a driving-related risk evaluation model based on the dynamic Bayesian network. The validity of this model is further tested by experimental data. The results show that this model can achieve a reasonable quantitative evaluation of driving-related risks. Vehicle operation-related risks can be further divided into four levels of safety, namely, levels I to IV. The lowest risk is observed at level I, whereas level IV has the highest risk. Among the indicators for risk evaluation, the distance between vehicles is the most sensitive control indicator of vehicle operation-related risks. The research findings provide various methodologies that could be utilized for evaluation and early warning of driving-related risks. Thus, a theoretical foundation to seek solutions for the safety of advanced driving assistance system is formed.