Kazushi WATANABE

Factor Analysis and Risk Assessment of Traffic Accidents on Winter Expressway

Kazushi SANO

In this study, factors affecting the occurrence of traffic accidents on expressways through Niigata Prefecture during the winter season were analyzed, and a model was constructed to estimate the probability of the occurrence of traffic accidents.
In the analysis of traffic accident occurrence factors, parameters were estimated using several generalized linear models for the created accident data set, and the appropriate model corresponding to the characteristics of the data was selected by conducting overdispersion and zero-excess tests. As a result of each test, a zero excess negative binomial regression model was adopted. The results of the parameter estimation revealed that average vehicle speed, the percentage of large vehicles, minimum temperature, 6-hour snowfall, maximum downhill gradient, dry road surface, and thin, frozen, or compacted snow were significant factors influencing the occurrence of traffic accidents.
In addition to the zero-excess negative binomial regression model, the lightGBM machine learning method was used to estimate the probability of traffic accidents. Both models were evaluated using five different patterns of training and test data. The results of the accident probability estimation showed that both models have the same estimation trend for all data, confirming that both models are general-purpose models that do not depend on a specific year. In addition, the false negatives tended to increase as the accuracy of the accident occurrence interval was increased. On the other hand, when the criteria for judgment were relaxed in order to reduce the number of missed accidents, false positives tended to increase.
Considering its practical application, lightGBM is effective for rapid risk assessment because it does not require the selection of explanatory variables for the model every time data changes and is computationally fast. On the other hand, the regression model is suitable for examining specific accident prevention measures because it can analyze factors that affect the probability of accidents based on the estimated coefficients. Therefore, it is necessary to select a model according to the purpose.