Modeling Towards Highway Traffic Crash Prediction Considering Spatial-Temporal Auto-Correlation

Through the analysis of road traffic accident data,abnormal driving behavior data,etc.,the highway traffic accident hotspots can be identified accurately,which can provide suggestions for safety improvement.However,the existing road traffic accident frequency modeling approaches lacks the multi-variate modeling of single-vehicle crashes and vehicle-to-vehicle crashes,and less consideration is given to the temporal auto-correlation and spatial autocorrelation between the variables,which leads to the deviation of the identification results from the actual situation.Therefore,this thesis studies the multivariate crash frequency modeling and hotspots identification method with considering of temporal and spatial autocorrelation.This thesis mainly completes the following aspects:(1)Modeling and analysis of crash frequency considering the spatial auto-correlation from cross-section dataBased on the cross-sectional traffic accident data,a Poisson distribution model,a multivariate spatial autocorrelation Poisson model and a multivariate spatial autocorrelation negative binomial model were established respectively.By the comparison of the logarithmic marginal prediction likelihood,Watanabe-Akaike Information criterion and the Deviance Information criterion,the multivariate spatial autocorrelation negative binomial distribution model is determined as the best model,indicating that the multivariate joint modeling with the consideration of spatial autocorrelation is better in model fitting performance.According to the parameter estimation results of the multivariate spatial autocorrelation negative binomial distribution model,the influencing factors of singlevehicle crashes and multi-vehicle crashes were analyzed.The results show that a significant correlation between single-vehicle crashes and vehicle-to-vehicle crashes,traffic volume,rigid guardrail Isolation facility,service area index,abnormal driving behavior variation index are significant,and are positively correlated with the number of two types of accidents.However,the number and average of abnormal driving behaviors of sudden acceleration and sudden left lane change only have a significant impact on the number of vehicle-to-vehicle crashes.(2)Modeling and analysis of crash frequency considering the spatial-temporal autocorrelation from panel dataFor the daily panel accident data and the monthly panel accident data,a first-order autoregressive multivariate spatiotemporal autocorrelation monthly crash count model,a first-order autoregressive multivariate spatiotemporal autocorrelation daily crash count model and a second-order autoregressive multivariate model were established respectively.The results show that the second-order autoregressive multivariate spatiotemporal autocorrelation daily crash count model has the best fit.The spatial autocorrelation coefficient was significant in all 3 spatial-temporal auto-correlation models,but the temporal autocorrelation coefficient was only significant in the two-day crash count model.The parameter estimation results based on the second-order autoregressive multivariate spatiotemporal autocorrelation model of daily crash counts show that,except for the number of abnormal driving behaviors,the mean of the severity of 2 abnormal driving behaviors,and the standard deviation of the severity of 2 abnormal driving behaviors,the significant parameters are basically consistent with the spatial autocorrelation model,and weather and wind have a significant impact on the number of single-vehicle crashes and the number of vehicle-to-vehicle crashes.(3)Hotspots identification considering spatiotemporal autocorrelationTaking the cross-sectional accident data as the object,through the method difference test and method consistency test,the empirical Bayesian potential safety improvement based on the spatiotemporal model and the absolute crash count method are compared.At the same time,the consistency test is used to compare the spatial difference between the single-vehicle crashes and multi-vehicle crashes hotspots.taking the panel data as the object,the multi-level consistency test is used to compare the first-order autoregression spatialtemporal autocorrelation model.The difference between the full Bayesian potential safety improvement index method and absolute crash-count based method in judging accident hotspots.The results show that for both cross-sectional accident data and daily panel accident data,the empirical Bayesian potential safety improvement index method better than absolute crash-count-based method in method consistency and method difference test.According to the number method,there is a big difference in the spatial distribution of the single-vehicle crashes and multi-vehicle crashes hotspots.