Accident Prediction Models For Freeways Based On Road Alignments

Highway geometry alignment is one of objective reasons for traffic accidents,especially in mountainous hilly areas. Combining practical scientific research and basedon the analysis on the relation between geometry alignments and accident indexes,prominent alignment factors are identified, and then the research about overall accidentprediction models and the number of casualties are carried out.Select north-Guangdong section of Jingzhu freeway, Guangdong section of Yueganfreeway and Kaiyang freeway as objects of study in mountains, in hills and in plains.Firstly, Road cells are divided based on geometry alignments in the way ofnon-fixed-length division. Secondly, the technique to determine OD matrix is improvedto obtain link traffic volumes according to volume data from toll booths. Thirdly, eachaccident is determined which road cell belonging to by mileage, so that associationrelationship between accidents and road cells is established. At last, databases ofaccidents, geometry alignments and the relation are built, for the purpose to manage thedata above uniformly and to retrieve and analyze data expediently.For the situation of variable selection, a combination of BP Neural Network andSensitivity Analysis is developed. In mountainous and rolling areas, three alignments oflongitudinal gradient(i), radius of horizontal curve(RHC) and length of tangent(LT) haveprominent influence on accidents happening. While geometry alignments have littleeffect on accidents in plains, which mainly depend on AADT. Further study demonstratethat, for freeways in mountains and hills, there is a twice correlation between i(or LT)and accident rate, and a negative correlation between RHC and accident rate(biggerradius, lower accident rate).On the foundation of identification of prominent factors, AADT is taken asindependent variable to model accident prediction model with base conditions, and thencorrection factors of prominent alignment factors are obtained to form accidentprediction system. Because geometry alignments have little effect on accidents forfreeways in plains, basic model is taken as the final result. The predicted results showthat the model can predict accident of each cell, errors of the total number of accidentsbetween the predicted value and observed value varies from-7.87%to4.90%.For the purpose of more comprehensive safety evaluation, severe degrees ofaccidents should be taken into consideration on the foundation of overall accidentprediction. This paper take the number of casualties of each cell as the index of decisionabout sever degrees. Because of large amount of zero-count data, to make modelselection more scientific, Vuong statistic, t statistic of the NB overdispersion parameterα, AIC and BIC are chosen to be distinguishing indexes, and Poisson regression modelsof NB、ZIP-Logit、ZINP-Probit are used to predict the number of injured and death toll. The forecast results have a high accuracy with an error of-0.2%~7.58%.