A Bayesian Network Model for Real-time Crash Prediction Based on Selected Variables by Random Forest

To reduce crash risk through managing real-time traffic flow is an effective way to improve highway safety. Populating crash data with the high-resolution (in space and time) traffic data to investigate the relationship between crash risk and traffic conditions can help the active traffic management (ATM) to predict crash risk in real-time. To improve the accuracy of the existing real-time crash prediction methods, this paper proposes a Bayesian network (BN) model with the variables selected by the random forest (RF). The study used the traffic and crash data from I-5 segment (13.8-miles-long) in Los Angeles, where the traffic data (flow, speed, occupancy and etc.) were recorded by sensors. The RF was used to rank the explanatory variables based on the Gini index, which yields the most significant variables for the crash prediction. Different from the previous studies, the gradient change of traffic data along distance is recognized as an important variable. The developed BN-RF model is evaluated by ROC curve. The results indicate that traffic conditions at two five-minute intervals prior to a crash is very sensitive to the crash risk prediction. A sudden speed change (speed magnitude in a short time and distance), characterized as one unit, makes the distribution of crash posterior probability at least 0.045 higher compared to the marginal value. The most important finding is that the proposed BN-RF model for the real-time crash prediction can accurately predict 70.46% of crashes with a 16.07% of false alarm rate, which is better than that of previous studies.